rebase on master after whisper_state changes

coreml : simlpify whisper_encode + log messages
coreml : use Core ML encoder inference
2025-06-24 17:15:19 +00:00 · 2023-03-26 16:09:06 +03:00 · 2023-03-26 15:48:45 +03:00 · 2023-03-26 15:48:41 +03:00
49 changed files with 1408 additions and 8324 deletions
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@ -265,44 +265,3 @@ jobs:
          popd
          emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
          make
-
-  ios:
-    runs-on: macos-latest
-
-    strategy:
-      matrix:
-        build: [Release]
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v1
-
-      - name: Configure
-        run: cp models/for-tests-ggml-base.en.bin models/ggml-base.en.bin
-
-      - name: Build objc example
-        run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build
-      
-      - name: Build swiftui example
-        run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphonesimulator build
-
-  android:
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v1
-
-      - name: Install Java
-        uses: actions/setup-java@v3
-        with:
-          distribution: zulu
-          java-version: 17
-      
-      - name: Setup Android SDK
-        uses: android-actions/setup-android@v2
-
-      - name: Build
-        run: |
-          cd examples/whisper.android
-          ./gradlew assembleRelease --no-daemon
--- a/.gitignore
+++ b/.gitignore
@ -1,8 +1,8 @@
 *.o
 *.a
+*.mlmodel
+*.mlmodelc
 .cache/
-.coreml/
-.test/
 .vs/
 .vscode/
 .DS_Store
@ -20,7 +20,6 @@ build-sanitize-thread/
 /stream
 /command
 /talk
-/talk-llama
 /bench

 arm_neon.h
@ -35,7 +34,3 @@ examples/whisper.objc/whisper.objc.xcodeproj/xcuserdata/
 examples/whisper.objc/whisper.objc.xcodeproj/project.xcworkspace/xcuserdata

 extra/bench-gg.txt
-
-models/*.mlmodel
-models/*.mlmodelc
-models/*.mlpackage
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -1,10 +1,6 @@
 cmake_minimum_required (VERSION 3.0)

-project(whisper.cpp VERSION 1.3.0)
-
-if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
-    add_compile_options(/utf-8)
-endif ()
+project(whisper.cpp VERSION 1.2.1)

 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
--- a/2
+++ b/2
@ -1,6 +1,6 @@
 MIT License

-Copyright (c) 2023 Georgi Gerganov
+Copyright (c) 2022 Georgi Gerganov

 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal
--- a/44
+++ b/44
@ -36,7 +36,7 @@ LDFLAGS  =

 # ref: https://github.com/ggerganov/whisper.cpp/issues/37
 ifneq ($(wildcard /usr/include/musl/*),)
-	CFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+	CFLAGS   += -D_POSIX_SOURCE -D_GNU_SOURCE
 	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
 endif

@ -77,6 +77,10 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 			CFLAGS += -mavx2
 		endif
 	else ifeq ($(UNAME_S),Linux)
+		AVX1_M := $(shell grep "avx " /proc/cpuinfo)
+		ifneq (,$(findstring avx,$(AVX1_M)))
+			CFLAGS += -mavx
+		endif
 		AVX2_M := $(shell grep "avx2 " /proc/cpuinfo)
 		ifneq (,$(findstring avx2,$(AVX2_M)))
 			CFLAGS += -mavx2
@ -88,17 +92,16 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 		F16C_M := $(shell grep "f16c " /proc/cpuinfo)
 		ifneq (,$(findstring f16c,$(F16C_M)))
 			CFLAGS += -mf16c
-
-			AVX1_M := $(shell grep "avx " /proc/cpuinfo)
-			ifneq (,$(findstring avx,$(AVX1_M)))
-				CFLAGS += -mavx
-			endif
 		endif
 		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
 		ifneq (,$(findstring sse3,$(SSE3_M)))
 			CFLAGS += -msse3
 		endif
 	else ifeq ($(UNAME_S),Haiku)
+		AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
+		ifneq (,$(findstring avx,$(AVX1_M)))
+			CFLAGS += -mavx
+		endif
 		AVX2_M := $(shell sysinfo -cpu | grep "AVX2 ")
 		ifneq (,$(findstring avx2,$(AVX2_M)))
 			CFLAGS += -mavx2
@ -110,11 +113,6 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 		F16C_M := $(shell sysinfo -cpu | grep "F16C ")
 		ifneq (,$(findstring f16c,$(F16C_M)))
 			CFLAGS += -mf16c
-
-			AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
-			ifneq (,$(findstring avx,$(AVX1_M)))
-				CFLAGS += -mavx
-			endif
 		endif
 	else
 		CFLAGS += -mfma -mf16c -mavx -mavx2
@ -157,15 +155,12 @@ ifneq ($(filter aarch64%,$(UNAME_M)),)
 	CXXFLAGS += -mcpu=native
 endif
 ifneq ($(filter armv6%,$(UNAME_M)),)
-	# 32-bit Raspberry Pi 1, 2, 3
-	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access
+	# Raspberry Pi 1, 2, 3
+	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
 endif
 ifneq ($(filter armv7%,$(UNAME_M)),)
-	# 32-bit ARM, for example on Armbian or possibly raspbian
-	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
-
-	# 64-bit ARM, use these (TODO: auto-detect 64-bit)
-	# CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	# Raspberry Pi 4
+	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 4
@ -187,7 +182,7 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )

-default: main bench
+default: main

 #
 # Build library
@ -196,7 +191,7 @@ default: main bench
 ggml.o: ggml.c ggml.h
 	$(CC)  $(CFLAGS)   -c ggml.c -o ggml.o

-whisper.o: whisper.cpp whisper.h ggml.h
+whisper.o: whisper.cpp whisper.h
 	$(CXX) $(CXXFLAGS) -c whisper.cpp -o whisper.o

 ifndef WHISPER_COREML
@ -218,7 +213,7 @@ libwhisper.so: ggml.o $(WHISPER_OBJ)
 	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)

 clean:
-	rm -f *.o main stream command talk talk-llama bench libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk bench libwhisper.a libwhisper.so

 #
 # Examples
@ -233,9 +228,6 @@ main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
 	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o main $(LDFLAGS)
 	./main -h

-bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
-
 stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
 	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)

@ -245,8 +237,8 @@ command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(W
 talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
 	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)

-talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
+bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)

 #
 # Audio samples
--- a/README.md
+++ b/README.md
@ -4,12 +4,12 @@
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)

-Beta: [v1.3.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.3.0) / Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)

 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:

 - Plain C/C++ implementation without dependencies
- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate framework and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
+- Apple silicon first-class citizen - optimized via Arm Neon and Accelerate framework
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
@ -58,9 +58,7 @@ the Accelerate framework utilizes the special-purpose AMX coprocessor available

 ## Quick start

-First clone the repository.
-
-Then, download one of the Whisper models converted in [ggml format](models). For example:
+First, download one of the Whisper models converted in [ggml format](models). For example:

 ```bash
 bash ./models/download-ggml-model.sh base.en
@ -225,60 +223,6 @@ make large
 | medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
 | large  | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |

-## Core ML support
-
-On Apple Silicon devices, the Encoder inference can be executed on the Apple Neural Engine (ANE) via Core ML. This can result in significant
-speed-up - more than x3 faster compared with CPU-only execution. Here are the instructions for generating a Core ML model and using it with `whisper.cpp`:
-
- Install Python dependencies needed for the creation of the Core ML model:
-
-  ```bash
-  pip install ane_transformers
-  pip install openai-whisper
-  pip install coremltools
-  ```
-
- Generate a Core ML model. For example, to generate a `base.en` model, use:
-
-  ```bash
-  ./models/generate-coreml-model.sh base.en
-  ```
-
-  This will generate the folder `models/ggml-base.en-encoder.mlmodelc`
-
- Build `whisper.cpp` with Core ML support:
-
-  ```bash
-  # using Makefile
-  make clean
-  WHISPER_COREML=1 make -j
-  
-  # using CMake
-  cd build
-  cmake -DWHISPER_COREML=1 ..
-  ```
-
- Run the examples as usual. For example:
-
-  ```bash
-  ./main -m models/ggml-base.en.bin -f samples/jfk.wav
-
-  ...
-
-  whisper_init_state: loading Core ML model from 'models/ggml-base.en-encoder.mlmodelc'
-  whisper_init_state: first run on a device may take a while ...
-  whisper_init_state: Core ML model loaded
-
-  system_info: n_threads = 4 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | COREML = 1 | 
-
-  ...
-  ```
-
-  The first run on a device is slow, since the ANE service compiles the Core ML model to some device-specific format.
-  Next runs are faster.
-  
-For more information about the Core ML implementation please refer to PR [#566](https://github.com/ggerganov/whisper.cpp/pull/566).
-  
 ## Limitations

 - Inference only
@ -369,7 +313,7 @@ whisper_print_timings:    total time = 32733.52 ms
 ## Real-time audio input example

 This is a naive example of performing real-time inference on audio from your microphone.
-The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continuously.
+The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).

 ```java
@ -384,10 +328,6 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Adding the `--print-colors` argument will print the transcribed text using an experimental color coding strategy
 to highlight words with high or low confidence:

-```java
-./main -m models/ggml-base.en.bin -f samples/gb0.wav --print-colors
-```
-
 <img width="965" alt="image" src="https://user-images.githubusercontent.com/1991296/197356445-311c8643-9397-4e5e-b46e-0b4b4daa2530.png">

 ## Controlling the length of the generated text segments (experimental)
@ -540,7 +480,6 @@ in [models](models).
 - [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
 - [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
 - [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
-  - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)
 - [X] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
  - [sandrohanea/whisper.net](https://github.com/sandrohanea/whisper.net)
  - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
@ -548,7 +487,6 @@ in [models](models).
  - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
  - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
 - [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
- [X] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)

 ## Examples

@ -562,7 +500,6 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 | [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
 | [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
 | [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
-| [talk-llama](examples/talk-llama) | | Talk with a LLaMA bot |
 | [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
 | [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
 | [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |
--- a/bindings/go/params.go
+++ b/bindings/go/params.go
@ -105,10 +105,6 @@ func (p *Params) SetMaxSegmentLength(n int) {
 	p.max_len = C.int(n)
 }

-func (p *Params) SetTokenTimestamps(b bool) {
-	p.token_timestamps = toBool(b)
-}
-
 // Set max tokens per segment (0 = no limit)
 func (p *Params) SetMaxTokensPerSegment(n int) {
 	p.max_tokens = C.int(n)
--- a/bindings/go/pkg/whisper/context.go
+++ b/bindings/go/pkg/whisper/context.go
@ -111,11 +111,6 @@ func (context *context) SetMaxSegmentLength(n uint) {
 	context.params.SetMaxSegmentLength(int(n))
 }

-// Set token timestamps flag
-func (context *context) SetTokenTimestamps(b bool) {
-	context.params.SetTokenTimestamps(b)
-}
-
 // Set max tokens per segment (0 = no limit)
 func (context *context) SetMaxTokensPerSegment(n uint) {
 	context.params.SetMaxTokensPerSegment(int(n))
@ -285,14 +280,10 @@ func toSegment(ctx *whisper.Context, n int) Segment {
 func toTokens(ctx *whisper.Context, n int) []Token {
 	result := make([]Token, ctx.Whisper_full_n_tokens(n))
 	for i := 0; i < len(result); i++ {
-		data := ctx.Whisper_full_get_token_data(n, i)
-
 		result[i] = Token{
-			Id:    int(ctx.Whisper_full_get_token_id(n, i)),
-			Text:  ctx.Whisper_full_get_token_text(n, i),
-			P:     ctx.Whisper_full_get_token_p(n, i),
-			Start: time.Duration(data.T0()) * time.Millisecond * 10,
-			End:   time.Duration(data.T1()) * time.Millisecond * 10,
+			Id:   int(ctx.Whisper_full_get_token_id(n, i)),
+			Text: strings.TrimSpace(ctx.Whisper_full_get_token_text(n, i)),
+			P:    ctx.Whisper_full_get_token_p(n, i),
 		}
 	}
 	return result
--- a/bindings/go/pkg/whisper/interface.go
+++ b/bindings/go/pkg/whisper/interface.go
@ -41,7 +41,6 @@ type Context interface {
 	SetTokenThreshold(float32)    // Set timestamp token probability threshold
 	SetTokenSumThreshold(float32) // Set timestamp token sum probability threshold
 	SetMaxSegmentLength(uint)     // Set max segment length in characters
-	SetTokenTimestamps(bool)      // Set token timestamps flag
 	SetMaxTokensPerSegment(uint)  // Set max tokens per segment (0 = no limit)

 	// Process mono audio data and return any errors.
@ -86,8 +85,7 @@ type Segment struct {

 // Token is a text or special token
 type Token struct {
-	Id         int
-	Text       string
-	P          float32
-	Start, End time.Duration
+	Id   int
+	Text string
+	P    float32
 }
--- a/bindings/go/whisper.go
+++ b/bindings/go/whisper.go
@ -356,7 +356,7 @@ func (ctx *Context) Whisper_full_get_token_id(segment int, token int) Token {

 // Get token data for the specified token in the specified segment.
 // This contains probabilities, timestamps, etc.
-func (ctx *Context) Whisper_full_get_token_data(segment int, token int) TokenData {
+func (ctx *Context) whisper_full_get_token_data(segment int, token int) TokenData {
 	return TokenData(C.whisper_full_get_token_data((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
 }

@ -407,11 +407,3 @@ func callEncoderBegin(user_data unsafe.Pointer) C.bool {
 	}
 	return true
 }
-
-func (t TokenData) T0() int64 {
-	return int64(t.t0)
-}
-
-func (t TokenData) T1() int64 {
-	return int64(t.t1)
-}
--- a/bindings/ios
+++ b/bindings/ios
--- a/bindings/javascript/package.json
+++ b/bindings/javascript/package.json
@ -1,6 +1,6 @@
 {
  "name": "whisper.cpp",
-  "version": "1.3.0",
+  "version": "1.2.1",
  "description": "Whisper speech recognition",
  "main": "whisper.js",
  "scripts": {
--- a/bindings/javascript/whisper.js
+++ b/bindings/javascript/whisper.js
--- a/coreml/whisper-decoder-impl.h
+++ b/coreml/whisper-decoder-impl.h
@ -1,146 +0,0 @@
-//
-// whisper-decoder-impl.h
-//
-// This file was automatically generated and should not be edited.
-//
-
-#import <Foundation/Foundation.h>
-#import <CoreML/CoreML.h>
-#include <stdint.h>
-#include <os/log.h>
-
-NS_ASSUME_NONNULL_BEGIN
-
-
-/// Model Prediction Input Type
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
-@interface whisper_decoder_implInput : NSObject<MLFeatureProvider>
-
-/// token_data as 1 by 1 matrix of 32-bit integers
-@property (readwrite, nonatomic, strong) MLMultiArray * token_data;
-
-/// audio_data as 1 × 384 × 1 × 1500 4-dimensional array of floats
-@property (readwrite, nonatomic, strong) MLMultiArray * audio_data;
- (instancetype)init NS_UNAVAILABLE;
- (instancetype)initWithToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data NS_DESIGNATED_INITIALIZER;
-
-@end
-
-
-/// Model Prediction Output Type
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
-@interface whisper_decoder_implOutput : NSObject<MLFeatureProvider>
-
-/// var_1346 as multidimensional array of floats
-@property (readwrite, nonatomic, strong) MLMultiArray * var_1346;
- (instancetype)init NS_UNAVAILABLE;
- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 NS_DESIGNATED_INITIALIZER;
-
-@end
-
-
-/// Class for model loading and prediction
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
-@interface whisper_decoder_impl : NSObject
-@property (readonly, nonatomic, nullable) MLModel * model;
-
-/**
-    URL of the underlying .mlmodelc directory.
-*/
-+ (nullable NSURL *)URLOfModelInThisBundle;
-
-/**
-    Initialize whisper_decoder_impl instance from an existing MLModel object.
-
-    Usually the application does not use this initializer unless it makes a subclass of whisper_decoder_impl.
-    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
-*/
- (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;
-
-/**
-    Initialize whisper_decoder_impl instance with the model in this bundle.
-*/
- (nullable instancetype)init;
-
-/**
-    Initialize whisper_decoder_impl instance with the model in this bundle.
-
-    @param configuration The model configuration object
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Initialize whisper_decoder_impl instance from the model URL.
-
-    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Initialize whisper_decoder_impl instance from the model URL.
-
-    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
-    @param configuration The model configuration object
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Construct whisper_decoder_impl instance asynchronously with configuration.
-    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
-
-    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
-*/
-+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
-
-/**
-    Construct whisper_decoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
-
-    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
-
-    @param modelURL The model URL.
-    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
-*/
-+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
-
-/**
-    Make a prediction using the standard interface
-    @param input an instance of whisper_decoder_implInput to predict from
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the prediction as whisper_decoder_implOutput
-*/
- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Make a prediction using the standard interface
-    @param input an instance of whisper_decoder_implInput to predict from
-    @param options prediction options
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the prediction as whisper_decoder_implOutput
-*/
- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Make a prediction using the convenience interface
-    @param token_data as 1 by 1 matrix of 32-bit integers:
-    @param audio_data as 1 × 384 × 1 × 1500 4-dimensional array of floats:
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the prediction as whisper_decoder_implOutput
-*/
- (nullable whisper_decoder_implOutput *)predictionFromToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-
-/**
-    Batch prediction
-    @param inputArray array of whisper_decoder_implInput instances to obtain predictions from
-    @param options prediction options
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the predictions as NSArray<whisper_decoder_implOutput *>
-*/
- (nullable NSArray<whisper_decoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_decoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
-@end
-
-NS_ASSUME_NONNULL_END
--- a/coreml/whisper-decoder-impl.m
+++ b/coreml/whisper-decoder-impl.m
@ -1,201 +0,0 @@
-//
-// whisper-decoder-impl.m
-//
-// This file was automatically generated and should not be edited.
-//
-
-#if !__has_feature(objc_arc)
-#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
-#endif
-
-#import "whisper-decoder-impl.h"
-
-@implementation whisper_decoder_implInput
-
- (instancetype)initWithToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data {
-    self = [super init];
-    if (self) {
-        _token_data = token_data;
-        _audio_data = audio_data;
-    }
-    return self;
-}
-
- (NSSet<NSString *> *)featureNames {
-    return [NSSet setWithArray:@[@"token_data", @"audio_data"]];
-}
-
- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
-    if ([featureName isEqualToString:@"token_data"]) {
-        return [MLFeatureValue featureValueWithMultiArray:self.token_data];
-    }
-    if ([featureName isEqualToString:@"audio_data"]) {
-        return [MLFeatureValue featureValueWithMultiArray:self.audio_data];
-    }
-    return nil;
-}
-
-@end
-
-@implementation whisper_decoder_implOutput
-
- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 {
-    self = [super init];
-    if (self) {
-        _var_1346 = var_1346;
-    }
-    return self;
-}
-
- (NSSet<NSString *> *)featureNames {
-    return [NSSet setWithArray:@[@"var_1346"]];
-}
-
- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
-    if ([featureName isEqualToString:@"var_1346"]) {
-        return [MLFeatureValue featureValueWithMultiArray:self.var_1346];
-    }
-    return nil;
-}
-
-@end
-
-@implementation whisper_decoder_impl
-
-
-/**
-    URL of the underlying .mlmodelc directory.
-*/
-+ (nullable NSURL *)URLOfModelInThisBundle {
-    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"whisper_decoder_impl" ofType:@"mlmodelc"];
-    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load whisper-decoder-impl.mlmodelc in the bundle resource"); return nil; }
-    return [NSURL fileURLWithPath:assetPath];
-}
-
-
-/**
-    Initialize whisper_decoder_impl instance from an existing MLModel object.
-
-    Usually the application does not use this initializer unless it makes a subclass of whisper_decoder_impl.
-    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
-*/
- (instancetype)initWithMLModel:(MLModel *)model {
-    self = [super init];
-    if (!self) { return nil; }
-    _model = model;
-    if (_model == nil) { return nil; }
-    return self;
-}
-
-
-/**
-    Initialize whisper_decoder_impl instance with the model in this bundle.
-*/
- (nullable instancetype)init {
-    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle error:nil];
-}
-
-
-/**
-    Initialize whisper_decoder_impl instance with the model in this bundle.
-
-    @param configuration The model configuration object
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle configuration:configuration error:error];
-}
-
-
-/**
-    Initialize whisper_decoder_impl instance from the model URL.
-
-    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    MLModel *model = [MLModel modelWithContentsOfURL:modelURL error:error];
-    if (model == nil) { return nil; }
-    return [self initWithMLModel:model];
-}
-
-
-/**
-    Initialize whisper_decoder_impl instance from the model URL.
-
-    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
-    @param configuration The model configuration object
-    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-*/
- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    MLModel *model = [MLModel modelWithContentsOfURL:modelURL configuration:configuration error:error];
-    if (model == nil) { return nil; }
-    return [self initWithMLModel:model];
-}
-
-
-/**
-    Construct whisper_decoder_impl instance asynchronously with configuration.
-    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
-
-    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
-*/
-+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler {
-    [self loadContentsOfURL:(NSURL * _Nonnull)[self URLOfModelInThisBundle]
-              configuration:configuration
-          completionHandler:handler];
-}
-
-
-/**
-    Construct whisper_decoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
-
-    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
-
-    @param modelURL The model URL.
-    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
-*/
-+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler {
-    [MLModel loadContentsOfURL:modelURL
-                 configuration:configuration
-             completionHandler:^(MLModel *model, NSError *error) {
-        if (model != nil) {
-            whisper_decoder_impl *typedModel = [[whisper_decoder_impl alloc] initWithMLModel:model];
-            handler(typedModel, nil);
-        } else {
-            handler(nil, error);
-        }
-    }];
-}
-
- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    return [self predictionFromFeatures:input options:[[MLPredictionOptions alloc] init] error:error];
-}
-
- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
-    if (!outFeatures) { return nil; }
-    return [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[outFeatures featureValueForName:@"var_1346"].multiArrayValue];
-}
-
- (nullable whisper_decoder_implOutput *)predictionFromToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    whisper_decoder_implInput *input_ = [[whisper_decoder_implInput alloc] initWithToken_data:token_data audio_data:audio_data];
-    return [self predictionFromFeatures:input_ error:error];
-}
-
- (nullable NSArray<whisper_decoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_decoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    id<MLBatchProvider> inBatch = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray:inputArray];
-    id<MLBatchProvider> outBatch = [self.model predictionsFromBatch:inBatch options:options error:error];
-    if (!outBatch) { return nil; }
-    NSMutableArray<whisper_decoder_implOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
-    for (NSInteger i = 0; i < outBatch.count; i++) {
-        id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
-        whisper_decoder_implOutput * result = [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[resultProvider featureValueForName:@"var_1346"].multiArrayValue];
-        [results addObject:result];
-    }
-    return results;
-}
-
-@end
--- a/coreml/whisper-encoder-impl.h
+++ b/coreml/whisper-encoder-impl.h
@ -1,5 +1,5 @@
 //
-// whisper-encoder-impl.h
+// CoremlEncoder.h
 //
 // This file was automatically generated and should not be edited.
 //
@ -13,20 +13,20 @@ NS_ASSUME_NONNULL_BEGIN


 /// Model Prediction Input Type
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
-@interface whisper_encoder_implInput : NSObject<MLFeatureProvider>
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
+@interface CoremlEncoderInput : NSObject<MLFeatureProvider>

-/// logmel_data as 1 × 80 × 3000 3-dimensional array of floats
-@property (readwrite, nonatomic, strong) MLMultiArray * logmel_data;
+/// melSegment as 1 × 80 × 3000 3-dimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * melSegment;
 - (instancetype)init NS_UNAVAILABLE;
- (instancetype)initWithLogmel_data:(MLMultiArray *)logmel_data NS_DESIGNATED_INITIALIZER;
+- (instancetype)initWithMelSegment:(MLMultiArray *)melSegment NS_DESIGNATED_INITIALIZER;

@end


 /// Model Prediction Output Type
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
-@interface whisper_encoder_implOutput : NSObject<MLFeatureProvider>
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
+@interface CoremlEncoderOutput : NSObject<MLFeatureProvider>

 /// output as multidimensional array of floats
@property (readwrite, nonatomic, strong) MLMultiArray * output;
@ -37,8 +37,8 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v


 /// Class for model loading and prediction
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
-@interface whisper_encoder_impl : NSObject
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
+@interface CoremlEncoder : NSObject
@property (readonly, nonatomic, nullable) MLModel * model;

 /**
@ -47,20 +47,20 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
 + (nullable NSURL *)URLOfModelInThisBundle;

 /**
-    Initialize whisper_encoder_impl instance from an existing MLModel object.
+    Initialize CoremlEncoder instance from an existing MLModel object.

-    Usually the application does not use this initializer unless it makes a subclass of whisper_encoder_impl.
+    Usually the application does not use this initializer unless it makes a subclass of CoremlEncoder.
    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
 */
 - (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;

 /**
-    Initialize whisper_encoder_impl instance with the model in this bundle.
+    Initialize CoremlEncoder instance with the model in this bundle.
 */
 - (nullable instancetype)init;

 /**
-    Initialize whisper_encoder_impl instance with the model in this bundle.
+    Initialize CoremlEncoder instance with the model in this bundle.

    @param configuration The model configuration object
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
@ -68,75 +68,75 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
 - (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;

 /**
-    Initialize whisper_encoder_impl instance from the model URL.
+    Initialize CoremlEncoder instance from the model URL.

-    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
 */
 - (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;

 /**
-    Initialize whisper_encoder_impl instance from the model URL.
+    Initialize CoremlEncoder instance from the model URL.

-    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
    @param configuration The model configuration object
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
 */
 - (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;

 /**
-    Construct whisper_encoder_impl instance asynchronously with configuration.
+    Construct CoremlEncoder instance asynchronously with configuration.
    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.

    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
 */
-+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler API_AVAILABLE(macos(11.0), ios(14.0), watchos(7.0), tvos(14.0)) __attribute__((visibility("hidden")));

 /**
-    Construct whisper_encoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+    Construct CoremlEncoder instance asynchronously with URL of .mlmodelc directory and optional configuration.

    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.

    @param modelURL The model URL.
    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
 */
-+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler API_AVAILABLE(macos(11.0), ios(14.0), watchos(7.0), tvos(14.0)) __attribute__((visibility("hidden")));

 /**
    Make a prediction using the standard interface
-    @param input an instance of whisper_encoder_implInput to predict from
+    @param input an instance of CoremlEncoderInput to predict from
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the prediction as whisper_encoder_implOutput
+    @return the prediction as CoremlEncoderOutput
 */
- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;

 /**
    Make a prediction using the standard interface
-    @param input an instance of whisper_encoder_implInput to predict from
+    @param input an instance of CoremlEncoderInput to predict from
    @param options prediction options
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the prediction as whisper_encoder_implOutput
+    @return the prediction as CoremlEncoderOutput
 */
- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;

 /**
    Make a prediction using the convenience interface
-    @param logmel_data as 1 × 80 × 3000 3-dimensional array of floats:
+    @param melSegment as 1 × 80 × 3000 3-dimensional array of floats:
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the prediction as whisper_encoder_implOutput
+    @return the prediction as CoremlEncoderOutput
 */
- (nullable whisper_encoder_implOutput *)predictionFromLogmel_data:(MLMultiArray *)logmel_data error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+- (nullable CoremlEncoderOutput *)predictionFromMelSegment:(MLMultiArray *)melSegment error:(NSError * _Nullable __autoreleasing * _Nullable)error;

 /**
    Batch prediction
-    @param inputArray array of whisper_encoder_implInput instances to obtain predictions from
+    @param inputArray array of CoremlEncoderInput instances to obtain predictions from
    @param options prediction options
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
-    @return the predictions as NSArray<whisper_encoder_implOutput *>
+    @return the predictions as NSArray<CoremlEncoderOutput *>
 */
- (nullable NSArray<whisper_encoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_encoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+- (nullable NSArray<CoremlEncoderOutput *> *)predictionsFromInputs:(NSArray<CoremlEncoderInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
@end

 NS_ASSUME_NONNULL_END
--- a/coreml/whisper-encoder-impl.m
+++ b/coreml/whisper-encoder-impl.m
@ -1,5 +1,5 @@
 //
-// whisper-encoder-impl.m
+// CoremlEncoder.m
 //
 // This file was automatically generated and should not be edited.
 //
@ -10,30 +10,30 @@

 #import "whisper-encoder-impl.h"

-@implementation whisper_encoder_implInput
+@implementation CoremlEncoderInput

- (instancetype)initWithLogmel_data:(MLMultiArray *)logmel_data {
+- (instancetype)initWithMelSegment:(MLMultiArray *)melSegment {
    self = [super init];
    if (self) {
-        _logmel_data = logmel_data;
+        _melSegment = melSegment;
    }
    return self;
 }

 - (NSSet<NSString *> *)featureNames {
-    return [NSSet setWithArray:@[@"logmel_data"]];
+    return [NSSet setWithArray:@[@"melSegment"]];
 }

 - (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
-    if ([featureName isEqualToString:@"logmel_data"]) {
-        return [MLFeatureValue featureValueWithMultiArray:self.logmel_data];
+    if ([featureName isEqualToString:@"melSegment"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.melSegment];
    }
    return nil;
 }

@end

-@implementation whisper_encoder_implOutput
+@implementation CoremlEncoderOutput

 - (instancetype)initWithOutput:(MLMultiArray *)output {
    self = [super init];
@ -56,23 +56,23 @@

@end

-@implementation whisper_encoder_impl
+@implementation CoremlEncoder


 /**
    URL of the underlying .mlmodelc directory.
 */
 + (nullable NSURL *)URLOfModelInThisBundle {
-    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"whisper_encoder_impl" ofType:@"mlmodelc"];
-    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load whisper-encoder-impl.mlmodelc in the bundle resource"); return nil; }
+    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"CoremlEncoder" ofType:@"mlmodelc"];
+    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load CoremlEncoder.mlmodelc in the bundle resource"); return nil; }
    return [NSURL fileURLWithPath:assetPath];
 }


 /**
-    Initialize whisper_encoder_impl instance from an existing MLModel object.
+    Initialize CoremlEncoder instance from an existing MLModel object.

-    Usually the application does not use this initializer unless it makes a subclass of whisper_encoder_impl.
+    Usually the application does not use this initializer unless it makes a subclass of CoremlEncoder.
    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
 */
 - (instancetype)initWithMLModel:(MLModel *)model {
@ -85,7 +85,7 @@


 /**
-    Initialize whisper_encoder_impl instance with the model in this bundle.
+    Initialize CoremlEncoder instance with the model in this bundle.
 */
 - (nullable instancetype)init {
    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle error:nil];
@ -93,7 +93,7 @@


 /**
-    Initialize whisper_encoder_impl instance with the model in this bundle.
+    Initialize CoremlEncoder instance with the model in this bundle.

    @param configuration The model configuration object
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
@ -104,9 +104,9 @@


 /**
-    Initialize whisper_encoder_impl instance from the model URL.
+    Initialize CoremlEncoder instance from the model URL.

-    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
 */
 - (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error {
@ -117,9 +117,9 @@


 /**
-    Initialize whisper_encoder_impl instance from the model URL.
+    Initialize CoremlEncoder instance from the model URL.

-    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
    @param configuration The model configuration object
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
 */
@ -131,13 +131,13 @@


 /**
-    Construct whisper_encoder_impl instance asynchronously with configuration.
+    Construct CoremlEncoder instance asynchronously with configuration.
    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.

    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
 */
-+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler {
+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler {
    [self loadContentsOfURL:(NSURL * _Nonnull)[self URLOfModelInThisBundle]
              configuration:configuration
          completionHandler:handler];
@ -145,20 +145,20 @@


 /**
-    Construct whisper_encoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+    Construct CoremlEncoder instance asynchronously with URL of .mlmodelc directory and optional configuration.

    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.

    @param modelURL The model URL.
    @param configuration The model configuration
-    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
 */
-+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler {
+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler {
    [MLModel loadContentsOfURL:modelURL
                 configuration:configuration
             completionHandler:^(MLModel *model, NSError *error) {
        if (model != nil) {
-            whisper_encoder_impl *typedModel = [[whisper_encoder_impl alloc] initWithMLModel:model];
+            CoremlEncoder *typedModel = [[CoremlEncoder alloc] initWithMLModel:model];
            handler(typedModel, nil);
        } else {
            handler(nil, error);
@ -166,29 +166,29 @@
    }];
 }

- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
    return [self predictionFromFeatures:input options:[[MLPredictionOptions alloc] init] error:error];
 }

- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
    id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
    if (!outFeatures) { return nil; }
-    return [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[outFeatures featureValueForName:@"output"].multiArrayValue];
+    return [[CoremlEncoderOutput alloc] initWithOutput:(MLMultiArray *)[outFeatures featureValueForName:@"output"].multiArrayValue];
 }

- (nullable whisper_encoder_implOutput *)predictionFromLogmel_data:(MLMultiArray *)logmel_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
-    whisper_encoder_implInput *input_ = [[whisper_encoder_implInput alloc] initWithLogmel_data:logmel_data];
+- (nullable CoremlEncoderOutput *)predictionFromMelSegment:(MLMultiArray *)melSegment error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    CoremlEncoderInput *input_ = [[CoremlEncoderInput alloc] initWithMelSegment:melSegment];
    return [self predictionFromFeatures:input_ error:error];
 }

- (nullable NSArray<whisper_encoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_encoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+- (nullable NSArray<CoremlEncoderOutput *> *)predictionsFromInputs:(NSArray<CoremlEncoderInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
    id<MLBatchProvider> inBatch = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray:inputArray];
    id<MLBatchProvider> outBatch = [self.model predictionsFromBatch:inBatch options:options error:error];
    if (!outBatch) { return nil; }
-    NSMutableArray<whisper_encoder_implOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
+    NSMutableArray<CoremlEncoderOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
    for (NSInteger i = 0; i < outBatch.count; i++) {
        id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
-        whisper_encoder_implOutput * result = [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[resultProvider featureValueForName:@"output"].multiArrayValue];
+        CoremlEncoderOutput * result = [[CoremlEncoderOutput alloc] initWithOutput:(MLMultiArray *)[resultProvider featureValueForName:@"output"].multiArrayValue];
        [results addObject:result];
    }
    return results;
--- a/coreml/whisper-encoder.mm
+++ b/coreml/whisper-encoder.mm
@ -18,7 +18,7 @@ struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {

    NSURL * url_model = [NSURL fileURLWithPath: path_model_str];

-    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model error:nil]);
+    const void * data = CFBridgingRetain([[CoremlEncoder alloc] initWithContentsOfURL:url_model error:nil]);

    if (data == NULL) {
        return NULL;
@ -49,16 +49,10 @@ void whisper_coreml_encode(
                                           error: nil
    ];

-    whisper_encoder_implOutput * outCoreML = [(__bridge id) ctx->data predictionFromLogmel_data:inMultiArray error:nil];
+    CoremlEncoderOutput * outCoreML = [(__bridge id) ctx->data predictionFromMelSegment:inMultiArray error:nil];

    MLMultiArray * outMA = outCoreML.output;

-    //NSArray<NSNumber *> * shape = outMA.shape;
-    //NSArray<NSNumber *> * strides = outMA.strides;
-
-    //printf("shape:   %ld %ld %ld %ld\n", [shape[0] longValue], [shape[1] longValue], [shape[2] longValue], [shape[3] longValue]);
-    //printf("strides: %ld %ld %ld %ld\n", [strides[0] longValue], [strides[1] longValue], [strides[2] longValue], [strides[3] longValue]);
-
    memcpy(out, outMA.dataPointer, outMA.count * sizeof(float));
 }

--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@ -63,5 +63,4 @@ else()
    add_subdirectory(command)
    add_subdirectory(bench)
    add_subdirectory(talk)
-    add_subdirectory(talk-llama)
 endif()
--- a/examples/addon.node/test/whisper.spec.js
+++ b/examples/addon.node/test/whisper.spec.js
@ -1,22 +1,15 @@
-const path = require("path");
-const { whisper } = require(path.join(
-  __dirname,
-  "../../../build/Release/whisper-addon"
-));
-const { promisify } = require("util");
-
-const whisperAsync = promisify(whisper);
+const path = require('path');
+const { whisper } = require(path.join(__dirname, '../../../build/Release/whisper-addon'));

 const whisperParamsMock = {
-  language: "en",
-  model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
-  fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
+    language: 'en',
+    model: path.join(__dirname, '../../../models/ggml-base.en.bin'),
+    fname_inp: path.join(__dirname, '../../../samples/jfk.wav'),
 };

 describe("Run whisper.node", () => {
-  test("it should receive a non-empty value", async () => {
-    let result = await whisperAsync(whisperParamsMock);

-    expect(result.length).toBeGreaterThan(0);
-  });
+    test("it should receive a non-empty value", () => {
+        expect(whisper(whisperParamsMock).length).toBeGreaterThan(0);
+    });
 });
--- a/examples/addon.node/addon.cpp
+++ b/examples/addon.node/addon.cpp
@ -160,6 +160,22 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
        return 3;
    }

+    // initial prompt
+    std::vector<whisper_token> prompt_tokens;
+
+    if (!params.prompt.empty()) {
+        prompt_tokens.resize(1024);
+        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
+
+        fprintf(stderr, "\n");
+        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
+        fprintf(stderr, "initial tokens: [ ");
+        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
+            fprintf(stderr, "%d ", prompt_tokens[i]);
+        }
+        fprintf(stderr, "]\n");
+    }
+
    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
        const auto fname_inp = params.fname_inp[f];
        const auto fname_out = f < (int)params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@ -227,7 +243,8 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
            wparams.greedy.best_of        = params.best_of;
            wparams.beam_search.beam_size = params.beam_size;

-            wparams.initial_prompt   = params.prompt.c_str();
+            wparams.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();

            whisper_print_user_data user_data = { &params, &pcmf32s };

--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@ -8,7 +8,6 @@
 #include <string>
 #include <thread>
 #include <vector>
-#include <cstring>

 // Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
 // Lowest is red, middle is yellow, highest is green.
@ -57,7 +56,7 @@ struct whisper_params {
    int32_t duration_ms  =  0;
    int32_t max_context  = -1;
    int32_t max_len      =  0;
-    int32_t best_of      =  2;
+    int32_t best_of      =  5;
    int32_t beam_size    = -1;

    float word_thold    =  0.01f;
@ -75,7 +74,6 @@ struct whisper_params {
    bool output_wts     = false;
    bool output_csv     = false;
    bool output_jsn     = false;
-    bool output_lrc     = false;
    bool print_special  = false;
    bool print_colors   = false;
    bool print_progress = false;
@ -131,7 +129,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-ovtt" || arg == "--output-vtt")     { params.output_vtt     = true; }
        else if (arg == "-osrt" || arg == "--output-srt")     { params.output_srt     = true; }
        else if (arg == "-owts" || arg == "--output-words")   { params.output_wts     = true; }
-        else if (arg == "-olrc" || arg == "--output-lrc")     { params.output_lrc     = true; }
        else if (arg == "-fp"   || arg == "--font-path")      { params.font_path      = argv[++i]; }
        else if (arg == "-ocsv" || arg == "--output-csv")     { params.output_csv     = true; }
        else if (arg == "-oj"   || arg == "--output-json")    { params.output_jsn     = true; }
@ -180,7 +177,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -otxt,     --output-txt        [%-7s] output result in a text file\n",                   params.output_txt ? "true" : "false");
    fprintf(stderr, "  -ovtt,     --output-vtt        [%-7s] output result in a vtt file\n",                    params.output_vtt ? "true" : "false");
    fprintf(stderr, "  -osrt,     --output-srt        [%-7s] output result in a srt file\n",                    params.output_srt ? "true" : "false");
-    fprintf(stderr, "  -olrc,     --output-lrc        [%-7s] output result in a lrc file\n",                    params.output_lrc ? "true" : "false");
    fprintf(stderr, "  -owts,     --output-words      [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
    fprintf(stderr, "  -fp,       --font-path         [%-7s] path to a monospace font for karaoke video\n",     params.font_path.c_str());
    fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
@ -211,8 +207,8 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper

    std::string speaker = "";

-    int64_t t0 = 0;
-    int64_t t1 = 0;
+    int64_t t0;
+    int64_t t1;

    // print the last n_new segments
    const int s0 = n_segments - n_new;
@ -375,39 +371,6 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
    return true;
 }

-char *escape_double_quotes(const char *str) {
-    if (str == NULL) {
-        return NULL;
-    }
-
-    size_t escaped_length = strlen(str) + 1;
-
-    for (size_t i = 0; str[i] != '\0'; i++) {
-        if (str[i] == '"') {
-            escaped_length++;
-        }
-    }
-
-    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
-    if (escaped == NULL) {
-        return NULL;
-    }
-
-    size_t pos = 0;
-    for (size_t i = 0; str[i] != '\0'; i++) {
-        if (str[i] == '"') {
-            escaped[pos++] = '\\';
-            escaped[pos++] = '"';
-        } else {
-            escaped[pos++] = str[i];
-        }
-    }
-
-    // no need to set zero due to calloc() being used prior
-
-    return escaped;
-}
-
 bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params) {
    std::ofstream fout(fname);
    int indent = 0;
@ -451,9 +414,7 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper

    auto value_s = [&](const char *name, const char *val, bool end = false) {
        start_value(name);
-        char * val_escaped = escape_double_quotes(val);
-        fout << "\"" << val_escaped << (end ? "\"\n" : "\",\n");
-        free(val_escaped);
+        fout << "\"" << val << (end ? "\"\n" : "\",\n");
    };

    auto end_value = [&](bool end = false) {
@ -494,7 +455,7 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
                value_i("ctx", whisper_model_n_text_ctx(ctx));
                value_i("state", whisper_model_n_text_state(ctx));
                value_i("head", whisper_model_n_text_head(ctx));
-                value_i("layer", whisper_model_n_text_layer(ctx), true);
+                value_i("leyer", whisper_model_n_text_layer(ctx), true);
            end_obj();
            value_i("mels", whisper_model_n_mels(ctx));
            value_i("f16", whisper_model_f16(ctx), true);
@ -516,7 +477,7 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);

                start_obj();
-                    start_obj("timestamps");
+                    start_obj("timestanps");
                        value_s("from", to_timestamp(t0, true).c_str());
                        value_s("to", to_timestamp(t1, true).c_str(), true);
                    end_obj();
@ -650,39 +611,6 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
    return true;
 }

-bool output_lrc(struct whisper_context * ctx, const char * fname) {
-
-    std::ofstream fout(fname);
-    if (!fout.is_open()) {
-        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
-        return false;
-    }
-
-    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
-
-    fout << "[by:whisper.cpp]\n";
-
-    const int n_segments = whisper_full_n_segments(ctx);
-    for (int i = 0; i < n_segments; ++i) {
-        const char * text = whisper_full_get_segment_text(ctx, i);
-        const int64_t t = whisper_full_get_segment_t0(ctx, i);
-
-        int64_t msec = t * 10;
-        int64_t min = msec / (1000 * 60);
-        msec = msec - min * (1000 * 60);
-        int64_t sec = msec / 1000;
-        msec = msec - sec * 1000;
-
-        char buf[16];
-        snprintf(buf, sizeof(buf), "%02d:%02d.%02d", (int) min, (int) sec, (int) ( msec / 10));
-        std::string timestamp_lrc = std::string(buf);
-
-        fout <<  '[' << timestamp_lrc << ']' << text << "\n";
-    }
-
-    return true;
-}
-
 int main(int argc, char ** argv) {
    whisper_params params;

@ -711,6 +639,22 @@ int main(int argc, char ** argv) {
        return 3;
    }

+    // initial prompt
+    std::vector<whisper_token> prompt_tokens;
+
+    if (!params.prompt.empty()) {
+        prompt_tokens.resize(1024);
+        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
+
+        fprintf(stderr, "\n");
+        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
+        fprintf(stderr, "initial tokens: [ ");
+        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
+            fprintf(stderr, "%d ", prompt_tokens[i]);
+        }
+        fprintf(stderr, "]\n");
+    }
+
    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
        const auto fname_inp = params.fname_inp[f];
 		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@ -774,7 +718,8 @@ int main(int argc, char ** argv) {

            wparams.speed_up         = params.speed_up;

-            wparams.initial_prompt   = params.prompt.c_str();
+            wparams.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();

            wparams.greedy.best_of        = params.best_of;
            wparams.beam_search.beam_size = params.beam_size;
@ -849,12 +794,6 @@ int main(int argc, char ** argv) {
                const auto fname_jsn = fname_out + ".json";
                output_json(ctx, fname_jsn.c_str(), params);
            }
-
-            // output to LRC file
-            if (params.output_lrc) {
-                const auto fname_lrc = fname_out + ".lrc";
-                output_lrc(ctx, fname_lrc.c_str());
-            }
        }
    }

--- a/examples/stream/stream.cpp
+++ b/examples/stream/stream.cpp
@ -43,7 +43,6 @@ struct whisper_params {

    bool speed_up      = false;
    bool translate     = false;
-    bool no_fallback   = false;
    bool print_special = false;
    bool no_context    = true;
    bool no_timestamps = false;
@ -74,7 +73,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
-        else if (arg == "-nf"  || arg == "--no-fallback")   { params.no_fallback   = true; }
        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
        else if (arg == "-kc"  || arg == "--keep-context")  { params.no_context    = false; }
        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
@ -96,23 +94,22 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "\n");
    fprintf(stderr, "options:\n");
    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n",    params.n_threads);
-    fprintf(stderr, "            --step N        [%-7d] audio step size in milliseconds\n",                params.step_ms);
-    fprintf(stderr, "            --length N      [%-7d] audio length in milliseconds\n",                   params.length_ms);
-    fprintf(stderr, "            --keep N        [%-7d] audio to keep from previous step in ms\n",         params.keep_ms);
-    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                              params.capture_id);
-    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",       params.max_tokens);
-    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
-    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",           params.vad_thold);
-    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                   params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
-    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
-    fprintf(stderr, "  -nf,      --no-fallback   [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
-    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
-    fprintf(stderr, "  -kc,      --keep-context  [%-7s] keep context between audio chunks\n",              params.no_context ? "false" : "true");
-    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                                params.language.c_str());
-    fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                     params.model.c_str());
-    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                          params.fname_out.c_str());
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "            --step N        [%-7d] audio step size in milliseconds\n",             params.step_ms);
+    fprintf(stderr, "            --length N      [%-7d] audio length in milliseconds\n",                params.length_ms);
+    fprintf(stderr, "            --keep N        [%-7d] audio to keep from previous step in ms\n",      params.keep_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -kc,      --keep-context  [%-7s] keep context between audio chunks\n",           params.no_context ? "false" : "true");
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                  params.model.c_str());
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
    fprintf(stderr, "\n");
 }

@ -151,7 +148,7 @@ int main(int argc, char ** argv) {

    // whisper init

-    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1){
+    if (whisper_lang_id(params.language.c_str()) == -1) {
        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
        whisper_print_usage(argc, argv, params);
        exit(0);
@ -300,8 +297,7 @@ int main(int argc, char ** argv) {
            wparams.speed_up         = params.speed_up;

            // disable temperature fallback
-            //wparams.temperature_inc  = -1.0f;
-            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
+            wparams.temperature_inc  = -1.0f;

            wparams.prompt_tokens    = params.no_context ? nullptr : prompt_tokens.data();
            wparams.prompt_n_tokens  = params.no_context ? 0       : prompt_tokens.size();
--- a/examples/talk-llama/.gitignore
+++ b/examples/talk-llama/.gitignore
@ -1 +0,0 @@
-audio.mp3
--- a/examples/talk-llama/CMakeLists.txt
+++ b/examples/talk-llama/CMakeLists.txt
@ -1,16 +0,0 @@
-if (WHISPER_SUPPORT_SDL2)
-    # talk-llama
-    set(TARGET talk-llama)
-    #add_executable(${TARGET} talk-llama.cpp llama.cpp)
-    #target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
-    #target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
-
-    # TODO: this is temporary
-    #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
-
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
-    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
-
-    include(DefaultTargetOptions)
-endif ()
--- a/examples/talk-llama/README.md
+++ b/examples/talk-llama/README.md
@ -1,36 +0,0 @@
-# talk-llama
-
-Talk with an LLaMA AI in your terminal
-
-[Demo Talk](https://user-images.githubusercontent.com/1991296/228024237-848f998c-c334-46a6-bef8-3271590da83b.mp4)
-
-## Building
-
-The `talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
-
-```bash
-# Install SDL2 on Linux
-sudo apt-get install libsdl2-dev
-
-# Install SDL2 on Mac OS
-brew install sdl2
-
-# Build the "talk-llama" executable
-make talk-llama
-
-# Run it
-./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
-```
-
- The `-mw` argument specifies the Whisper model that you would like to use. Recommended `base` or `small` for real-time experience
- The `-ml` argument specifies the LLaMA model that you would like to use. Read the instructions in https://github.com/ggerganov/llama.cpp for information about how to obtain a `ggml` compatible LLaMA model
-
-## TTS
-
-For best experience, this example needs a TTS tool to convert the generated text responses to voice.
-You can use any TTS engine that you would like - simply edit the [speak.sh](speak.sh) script to your needs.
-By default, it is configured to use MacOS's `say`, but you can use whatever you wish.
-
-## Discussion
-
-If you have any feedback, please let "us" know in the following discussion: https://github.com/ggerganov/whisper.cpp/discussions/672?converting=1
--- a/examples/talk-llama/eleven-labs.py
+++ b/examples/talk-llama/eleven-labs.py
@ -1,23 +0,0 @@
-import sys
-import importlib.util
-
-api_key = "" #Write your https://beta.elevenlabs.io api key here
-if not api_key:
-    print("To use elevenlabs you have to register to https://beta.elevenlabs.io and add your elevenlabs api key to examples/talk-llama/eleven-labs.py")
-    sys.exit()
-
-if importlib.util.find_spec("elevenlabs") is None:
-    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
-    sys.exit()
-
-from elevenlabs import ElevenLabs
-eleven = ElevenLabs(api_key)
-
-# Get a Voice object, by name or UUID
-voice = eleven.voices["Arnold"] #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
-
-# Generate the TTS
-audio = voice.generate(str(sys.argv[2:]))
-
-# Save the TTS to a file
-audio.save("audio") 
--- a/examples/talk-llama/llama.cpp
+++ b/examples/talk-llama/llama.cpp
--- a/examples/talk-llama/llama.h
+++ b/examples/talk-llama/llama.h
@ -1,173 +0,0 @@
-#ifndef LLAMA_H
-#define LLAMA_H
-
-#include <stddef.h>
-#include <stdint.h>
-#include <stdbool.h>
-
-#ifdef LLAMA_SHARED
-#    if defined(_WIN32) && !defined(__MINGW32__)
-#        ifdef LLAMA_BUILD
-#            define LLAMA_API __declspec(dllexport)
-#        else
-#            define LLAMA_API __declspec(dllimport)
-#        endif
-#    else
-#        define LLAMA_API __attribute__ ((visibility ("default")))
-#    endif
-#else
-#    define LLAMA_API
-#endif
-
-#define LLAMA_FILE_VERSION 1
-#define LLAMA_FILE_MAGIC 0x67676a74 // 'ggjt' in hex
-#define LLAMA_FILE_MAGIC_UNVERSIONED 0x67676d6c // pre-versioned files
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-    //
-    // C interface
-    //
-    // TODO: show sample usage
-    //
-
-    struct llama_context;
-
-    typedef int llama_token;
-
-    typedef struct llama_token_data {
-        llama_token id;  // token id
-
-        float p;     // probability of the token
-        float plog;  // log probability of the token
-
-    } llama_token_data;
-
-    typedef void (*llama_progress_callback)(float progress, void *ctx);
-
-    struct llama_context_params {
-        int n_ctx;   // text context
-        int n_parts; // -1 for default
-        int seed;    // RNG seed, 0 for random
-
-        bool f16_kv;     // use fp16 for KV cache
-        bool logits_all; // the llama_eval() call computes all logits, not just the last one
-        bool vocab_only; // only load the vocabulary, no weights
-        bool use_mmap;   // use mmap if possible
-        bool use_mlock;  // force system to keep model in RAM
-        bool embedding;  // embedding mode only
-
-        // called with a progress value between 0 and 1, pass NULL to disable
-        llama_progress_callback progress_callback;
-        // context pointer passed to the progress callback
-        void * progress_callback_user_data;
-    };
-
-    LLAMA_API struct llama_context_params llama_context_default_params();
-
-    LLAMA_API bool llama_mmap_supported();
-    LLAMA_API bool llama_mlock_supported();
-
-    // Various functions for loading a ggml llama model.
-    // Allocate (almost) all memory needed for the model.
-    // Return NULL on failure
-    LLAMA_API struct llama_context * llama_init_from_file(
-                             const char * path_model,
-            struct llama_context_params   params);
-
-    // Frees all allocated memory
-    LLAMA_API void llama_free(struct llama_context * ctx);
-
-    // TODO: not great API - very likely to change
-    // Returns 0 on success
-    LLAMA_API int llama_model_quantize(
-            const char * fname_inp,
-            const char * fname_out,
-                   int   itype);
-
-    // Returns the KV cache that will contain the context for the
-    // ongoing prediction with the model.
-    LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);
-
-    // Returns the size of the KV cache
-    LLAMA_API size_t llama_get_kv_cache_size(struct llama_context * ctx);
-
-    // Returns the number of tokens in the KV cache
-    LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
-
-    // Sets the KV cache containing the current context for the model
-    LLAMA_API void llama_set_kv_cache(
-            struct llama_context * ctx,
-                   const uint8_t * kv_cache,
-                          size_t   n_size,
-                             int   n_token_count);
-
-    // Run the llama inference to obtain the logits and probabilities for the next token.
-    // tokens + n_tokens is the provided batch of new tokens to process
-    // n_past is the number of tokens to use from previous eval calls
-    // Returns 0 on success
-    LLAMA_API int llama_eval(
-            struct llama_context * ctx,
-               const llama_token * tokens,
-                             int   n_tokens,
-                             int   n_past,
-                             int   n_threads);
-
-    // Convert the provided text into tokens.
-    // The tokens pointer must be large enough to hold the resulting tokens.
-    // Returns the number of tokens on success, no more than n_max_tokens
-    // Returns a negative number on failure - the number of tokens that would have been returned
-    // TODO: not sure if correct
-    LLAMA_API int llama_tokenize(
-            struct llama_context * ctx,
-                      const char * text,
-                     llama_token * tokens,
-                             int   n_max_tokens,
-                            bool   add_bos);
-
-    LLAMA_API int llama_n_vocab(struct llama_context * ctx);
-    LLAMA_API int llama_n_ctx  (struct llama_context * ctx);
-    LLAMA_API int llama_n_embd (struct llama_context * ctx);
-
-    // Token logits obtained from the last call to llama_eval()
-    // The logits for the last token are stored in the last row
-    // Can be mutated in order to change the probabilities of the next token
-    // Rows: n_tokens
-    // Cols: n_vocab
-    LLAMA_API float * llama_get_logits(struct llama_context * ctx);
-
-    // Get the embeddings for the input
-    // shape: [n_embd] (1-dimensional)
-    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
-
-    // Token Id -> String. Uses the vocabulary in the provided context
-    LLAMA_API const char * llama_token_to_str(struct llama_context * ctx, llama_token token);
-
-    // Special tokens
-    LLAMA_API llama_token llama_token_bos();
-    LLAMA_API llama_token llama_token_eos();
-
-    // TODO: improve the last_n_tokens interface ?
-    LLAMA_API llama_token llama_sample_top_p_top_k(
-       struct llama_context * ctx,
-          const llama_token * last_n_tokens_data,
-                        int   last_n_tokens_size,
-                        int   top_k,
-                      float   top_p,
-                      float   temp,
-                      float   repeat_penalty);
-
-    // Performance information
-    LLAMA_API void llama_print_timings(struct llama_context * ctx);
-    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
-
-    // Print system information
-    LLAMA_API const char * llama_print_system_info(void);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // LLAMA_H
--- a/examples/talk-llama/llama_internal.h
+++ b/examples/talk-llama/llama_internal.h
@ -1,12 +0,0 @@
-// Internal header to be included by llama.cpp and tests/benchmarks only.
-
-#ifndef LLAMA_INTERNAL_H
-#define LLAMA_INTERNAL_H
-
-#include <vector>
-#include <string>
-struct ggml_tensor;
-
-std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
-
-#endif // LLAMA_INTERNAL_H
--- a/examples/talk-llama/llama_util.h
+++ b/examples/talk-llama/llama_util.h
@ -1,383 +0,0 @@
-// Internal header to be included only by llama.cpp.
-// Contains wrappers around OS interfaces.
-
-#ifndef LLAMA_UTIL_H
-#define LLAMA_UTIL_H
-
-#include <cstdio>
-#include <cstdint>
-#include <cerrno>
-#include <cstring>
-#include <cstdarg>
-#include <cstdlib>
-#include <climits>
-
-#include <string>
-#include <vector>
-
-#ifdef __has_include
-    #if __has_include(<unistd.h>)
-        #include <unistd.h>
-        #if defined(_POSIX_MAPPED_FILES)
-            #include <sys/mman.h>
-        #endif
-    #endif
-#endif
-
-#if defined(_WIN32)
-    #define WIN32_LEAN_AND_MEAN
-    #define NOMINMAX
-    #include <windows.h>
-    #include <io.h>
-    #include <stdio.h> // for _fseeki64
-#endif
-
-#define LLAMA_ASSERT(x) \
-    do { \
-        if (!(x)) { \
-            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            abort(); \
-        } \
-    } while (0)
-
-#ifdef __GNUC__
-__attribute__((format(printf, 1, 2)))
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    LLAMA_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-};
-
-struct llama_file {
-    // use FILE * so we don't have to re-open the file to mmap
-    FILE * fp;
-    size_t size;
-
-    llama_file(const char * fname, const char * mode) {
-        fp = std::fopen(fname, mode);
-        if (fp == NULL) {
-            throw format("failed to open %s: %s", fname, std::strerror(errno));
-        }
-        seek(0, SEEK_END);
-        size = tell();
-        seek(0, SEEK_SET);
-    }
-
-    size_t tell() const {
-#ifdef _WIN32
-        __int64 ret = _ftelli64(fp);
-#else
-        long ret = std::ftell(fp);
-#endif
-        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
-        return (size_t) ret;
-    }
-
-    void seek(size_t offset, int whence) {
-#ifdef _WIN32
-        int ret = _fseeki64(fp, (__int64) offset, whence);
-#else
-        int ret = std::fseek(fp, (long) offset, whence);
-#endif
-        LLAMA_ASSERT(ret == 0); // same
-    }
-
-    void read_raw(void * ptr, size_t size) {
-        if (size == 0) {
-            return;
-        }
-        errno = 0;
-        std::size_t ret = std::fread(ptr, size, 1, fp);
-        if (ferror(fp)) {
-            throw format("read error: %s", strerror(errno));
-        }
-        if (ret != 1) {
-            throw std::string("unexpectedly reached end of file");
-        }
-    }
-
-    std::uint32_t read_u32() {
-        std::uint32_t ret;
-        read_raw(&ret, sizeof(ret));
-        return ret;
-    }
-
-    std::string read_string(std::uint32_t len) {
-        std::vector<char> chars(len);
-        read_raw(chars.data(), len);
-        return std::string(chars.data(), len);
-    }
-
-    void write_raw(const void * ptr, size_t size) {
-        if (size == 0) {
-            return;
-        }
-        errno = 0;
-        size_t ret = std::fwrite(ptr, size, 1, fp);
-        if (ret != 1) {
-            throw format("write error: %s", strerror(errno));
-        }
-    }
-
-    void write_u32(std::uint32_t val) {
-        write_raw(&val, sizeof(val));
-    }
-
-    ~llama_file() {
-        if (fp) {
-            std::fclose(fp);
-        }
-    }
-};
-
-#if defined(_WIN32)
-static std::string llama_format_win_err(DWORD err) {
-    LPSTR buf;
-    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
-                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
-    if (!size) {
-        return "FormatMessageA failed";
-    }
-    std::string ret(buf, size);
-    LocalFree(buf);
-    return ret;
-}
-#endif
-
-struct llama_mmap {
-    void * addr;
-    size_t size;
-
-    llama_mmap(const llama_mmap &) = delete;
-
-#ifdef _POSIX_MAPPED_FILES
-    static constexpr bool SUPPORTED = true;
-
-    llama_mmap(struct llama_file * file) {
-        size = file->size;
-        int fd = fileno(file->fp);
-        int flags = MAP_SHARED;
-#ifdef __linux__
-        flags |= MAP_POPULATE;
-#endif
-        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
-        close(fd);
-        if (addr == MAP_FAILED) {
-            throw format("mmap failed: %s", strerror(errno));
-        }
-
-        // Advise the kernel to preload the mapped memory
-        if (madvise(addr, file->size, MADV_WILLNEED)) {
-            fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
-                    strerror(errno));
-        }
-    }
-
-    ~llama_mmap() {
-        munmap(addr, size);
-    }
-#elif defined(_WIN32)
-    static constexpr bool SUPPORTED = true;
-
-    llama_mmap(struct llama_file * file) {
-        size = file->size;
-
-        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
-
-        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
-        DWORD error = GetLastError();
-        CloseHandle(hFile);
-
-        if (hMapping == NULL) {
-            throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());
-        }
-
-        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
-        error = GetLastError();
-        CloseHandle(hMapping);
-
-        if (addr == NULL) {
-            throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str());
-        }
-
-        // Advise the kernel to preload the mapped memory
-        WIN32_MEMORY_RANGE_ENTRY range;
-        range.VirtualAddress = addr;
-        range.NumberOfBytes = (SIZE_T)size;
-        if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
-            fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
-        }
-    }
-
-    ~llama_mmap() {
-        if (!UnmapViewOfFile(addr)) {
-            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
-        }
-    }
-#else
-    static constexpr bool SUPPORTED = false;
-
-    llama_mmap(struct llama_file *) {
-        throw std::string("mmap not supported");
-    }
-#endif
-};
-
-// Represents some region of memory being locked using mlock or VirtualLock;
-// will automatically unlock on destruction.
-struct llama_mlock {
-    void * addr = NULL;
-    size_t size = 0;
-    bool failed_already = false;
-
-    llama_mlock() {}
-    llama_mlock(const llama_mlock &) = delete;
-
-    ~llama_mlock() {
-        if (size) {
-            raw_unlock(addr, size);
-        }
-    }
-
-    void init(void * addr) {
-        LLAMA_ASSERT(this->addr == NULL && this->size == 0);
-        this->addr = addr;
-    }
-
-    void grow_to(size_t target_size) {
-        LLAMA_ASSERT(addr);
-        if (failed_already) {
-            return;
-        }
-        size_t granularity = lock_granularity();
-        target_size = (target_size + granularity - 1) & ~(granularity - 1);
-        if (target_size > size) {
-            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
-                size = target_size;
-            } else {
-                failed_already = true;
-            }
-        }
-    }
-
-#ifdef _POSIX_MEMLOCK_RANGE
-    static constexpr bool SUPPORTED = true;
-
-    size_t lock_granularity() {
-        return (size_t) sysconf(_SC_PAGESIZE);
-    }
-
-    #ifdef __APPLE__
-        #define MLOCK_SUGGESTION \
-            "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
-            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
-    #else
-        #define MLOCK_SUGGESTION \
-            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
-    #endif
-
-    bool raw_lock(const void * addr, size_t size) {
-        if (!mlock(addr, size)) {
-            return true;
-        } else {
-            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n" MLOCK_SUGGESTION,
-                    size, this->size, std::strerror(errno));
-            return false;
-        }
-    }
-
-    #undef MLOCK_SUGGESTION
-
-    void raw_unlock(void * addr, size_t size) {
-        if (munlock(addr, size)) {
-            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
-        }
-    }
-#elif defined(_WIN32)
-    static constexpr bool SUPPORTED = true;
-
-    size_t lock_granularity() {
-        SYSTEM_INFO si;
-        GetSystemInfo(&si);
-        return (size_t) si.dwPageSize;
-    }
-
-    bool raw_lock(void * addr, size_t size) {
-        for (int tries = 1; ; tries++) {
-            if (VirtualLock(addr, size)) {
-                return true;
-            }
-            if (tries == 2) {
-                fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
-                        size, this->size, llama_format_win_err(GetLastError()).c_str());
-                return false;
-            }
-
-            // It failed but this was only the first try; increase the working
-            // set size and try again.
-            SIZE_T min_ws_size, max_ws_size;
-            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
-                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
-                return false;
-            }
-            // Per MSDN: "The maximum number of pages that a process can lock
-            // is equal to the number of pages in its minimum working set minus
-            // a small overhead."
-            // Hopefully a megabyte is enough overhead:
-            size_t increment = size + 1048576;
-            // The minimum must be <= the maximum, so we need to increase both:
-            min_ws_size += size;
-            max_ws_size += size;
-            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
-                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
-                        llama_format_win_err(GetLastError()).c_str());
-                return false;
-            }
-        }
-    }
-
-    void raw_unlock(void * addr, size_t size) {
-        if (!VirtualUnlock(addr, size)) {
-            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
-                    llama_format_win_err(GetLastError()).c_str());
-        }
-    }
-#else
-    static constexpr bool SUPPORTED = false;
-
-    void raw_lock(const void * addr, size_t size) {
-        fprintf(stderr, "warning: mlock not supported on this system\n");
-    }
-
-    void raw_unlock(const void * addr, size_t size) {}
-#endif
-};
-
-// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
-struct llama_buffer {
-    uint8_t * addr = NULL;
-    size_t size = 0;
-
-    void resize(size_t size) {
-        delete[] addr;
-        addr = new uint8_t[size];
-        this->size = size;
-    }
-
-    ~llama_buffer() {
-        delete[] addr;
-    }
-};
-#endif
--- a/examples/talk-llama/prompts/talk-alpaca.txt
+++ b/examples/talk-llama/prompts/talk-alpaca.txt
@ -1,23 +0,0 @@
-Below is an instruction that describes a task. Write a response that appropriately completes the request.
-
-### Instruction:
-
-Write a text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
-{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
-There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
-The transcript only includes text, it does not include markup like HTML and Markdown.
-{1} responds with short and concise answers.
-
-### Response:
-
-{0}{4} Hello, {1}!
-{1}{4} Hello {0}! How may I help you today?
-{0}{4} What time is it?
-{1}{4} It is {2} o'clock.
-{0}{4} What year is it?
-{1}{4} We are in {3}.
-{0}{4} What is a cat?
-{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
-{0}{4} Name a color.
-{1}{4} Blue
-{0}{4}
--- a/examples/talk-llama/speak.sh
+++ b/examples/talk-llama/speak.sh
@ -1,21 +0,0 @@
-#!/bin/bash
-
-# Usage:
-#  speak.sh <voice_id> <text-to-speak>
-
-# espeak
-# Mac OS: brew install espeak
-# Linux: apt-get install espeak
-#
-#espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 "$2"
-
-# for Mac
-say "$2"
-
-# Eleven Labs
-# To use it, install the elevenlabs module from pip (pip install elevenlabs), register to https://beta.elevenlabs.io to get an api key and paste it in /examples/talk-llama/eleven-labs.py 
-#
-#wd=$(dirname $0)
-#script=$wd/eleven-labs.py
-#python3 $script $1 "$2" >/dev/null 2>&1
-#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1
--- a/examples/talk-llama/talk-llama.cpp
+++ b/examples/talk-llama/talk-llama.cpp
@ -1,550 +0,0 @@
-// Talk with AI
-//
-
-#include "common.h"
-#include "common-sdl.h"
-#include "whisper.h"
-#include "llama.h"
-
-#include <cassert>
-#include <cstdio>
-#include <fstream>
-#include <regex>
-#include <string>
-#include <thread>
-#include <vector>
-#include <regex>
-
-std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
-    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
-    std::vector<llama_token> res(text.size() + (int)add_bos);
-    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
-    assert(n >= 0);
-    res.resize(n);
-
-    return res;
-}
-
-// command-line parameters
-struct whisper_params {
-    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t voice_ms   = 10000;
-    int32_t capture_id = -1;
-    int32_t max_tokens = 32;
-    int32_t audio_ctx  = 0;
-
-    int32_t n_parts_llama = -1;
-
-    float vad_thold    = 0.6f;
-    float freq_thold   = 100.0f;
-
-    bool speed_up      = false;
-    bool translate     = false;
-    bool print_special = false;
-    bool print_energy  = false;
-    bool no_timestamps = true;
-    bool verbose_prompt = false;
-
-    std::string person      = "Georgi";
-    std::string language    = "en";
-    std::string model_wsp   = "models/ggml-base.en.bin";
-    std::string model_llama = "models/ggml-llama-7B.bin";
-    std::string speak       = "./examples/talk-llama/speak.sh";
-    std::string prompt      = "";
-    std::string fname_out;
-};
-
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
-
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
-    for (int i = 1; i < argc; i++) {
-        std::string arg = argv[i];
-
-        if (arg == "-h" || arg == "--help") {
-            whisper_print_usage(argc, argv, params);
-            exit(0);
-        }
-        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
-        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms      = std::stoi(argv[++i]); }
-        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
-        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
-        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
-        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
-        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "--n-parts-llama")                  { params.n_parts_llama = std::stoi(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
-        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
-        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
-        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
-        else if (arg == "--verbose-prompt")                 { params.verbose_prompt = true; }
-        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
-        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
-        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
-        else if (arg == "-ml"  || arg == "--model-llama")   { params.model_llama   = argv[++i]; }
-        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
-        else if (arg == "--prompt-file")                    {
-            std::ifstream file(argv[++i]);
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
-            if (params.prompt.back() == '\n') {
-                params.prompt.pop_back();
-            }
-        }
-        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
-        else {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            whisper_print_usage(argc, argv, params);
-            exit(0);
-        }
-    }
-
-    return true;
-}
-
-void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
-    fprintf(stderr, "\n");
-    fprintf(stderr, "usage: %s [options]\n", argv[0]);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
-    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
-    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
-    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
-    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
-    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
-    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
-    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
-    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
-    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
-    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
-    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
-    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
-    fprintf(stderr, "  -ml FILE, --model-llama   [%-7s] llama model file\n",                            params.model_llama.c_str());
-    fprintf(stderr, "  --n-parts-llama N         [%-7d] num parts in llama model file\n",               params.n_parts_llama);
-    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
-    fprintf(stderr, "  --prompt-file FNAME       [%-7s] file with custom prompt to start dialog\n",     "");
-    fprintf(stderr, "  --verbose-prompt          [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
-    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
-    fprintf(stderr, "\n");
-}
-
-std::string transcribe(
-        whisper_context * ctx,
-        const whisper_params & params,
-        const std::vector<float> & pcmf32,
-        const std::string prompt_text,
-        float & prob,
-        int64_t & t_ms) {
-    const auto t_start = std::chrono::high_resolution_clock::now();
-
-    prob = 0.0f;
-    t_ms = 0;
-
-    std::vector<whisper_token> prompt_tokens;
-
-    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
-
-    prompt_tokens.resize(1024);
-    prompt_tokens.resize(whisper_tokenize(ctx, prompt_text.c_str(), prompt_tokens.data(), prompt_tokens.size()));
-
-    wparams.print_progress   = false;
-    wparams.print_special    = params.print_special;
-    wparams.print_realtime   = false;
-    wparams.print_timestamps = !params.no_timestamps;
-    wparams.translate        = params.translate;
-    wparams.no_context       = true;
-    wparams.single_segment   = true;
-    wparams.max_tokens       = params.max_tokens;
-    wparams.language         = params.language.c_str();
-    wparams.n_threads        = params.n_threads;
-
-    wparams.prompt_tokens    = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
-    wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();
-
-    wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
-
-    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
-        return "";
-    }
-
-    int prob_n = 0;
-    std::string result;
-
-    const int n_segments = whisper_full_n_segments(ctx);
-    for (int i = 0; i < n_segments; ++i) {
-        const char * text = whisper_full_get_segment_text(ctx, i);
-
-        result += text;
-
-        const int n_tokens = whisper_full_n_tokens(ctx, i);
-        for (int j = 0; j < n_tokens; ++j) {
-            const auto token = whisper_full_get_token_data(ctx, i, j);
-
-            prob += token.p;
-            ++prob_n;
-        }
-    }
-
-    if (prob_n > 0) {
-        prob /= prob_n;
-    }
-
-    const auto t_end = std::chrono::high_resolution_clock::now();
-    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
-
-    return result;
-}
-
-const std::string k_prompt_whisper = R"(A conversation with a person called {1}.)";
-
-const std::string k_prompt_llama = R"(Text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
-{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
-There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
-The transcript only includes text, it does not include markup like HTML and Markdown.
-{1} responds with short and concise answers.
-
-{0}{4} Hello, {1}!
-{1}{4} Hello {0}! How may I help you today?
-{0}{4} What time is it?
-{1}{4} It is {2} o'clock.
-{0}{4} What year is it?
-{1}{4} We are in {3}.
-{0}{4} What is a cat?
-{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
-{0}{4} Name a color.
-{1}{4} Blue
-{0}{4})";
-
-int main(int argc, char ** argv) {
-    whisper_params params;
-
-    if (whisper_params_parse(argc, argv, params) == false) {
-        return 1;
-    }
-
-    if (whisper_lang_id(params.language.c_str()) == -1) {
-        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
-        whisper_print_usage(argc, argv, params);
-        exit(0);
-    }
-
-    // whisper init
-
-    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
-
-    // llama init
-
-    auto lparams = llama_context_default_params();
-
-    // tune these to your liking
-    lparams.n_ctx      = 2048;
-    lparams.seed       = 1;
-    lparams.f16_kv     = true;
-    lparams.n_parts    = params.n_parts_llama;
-
-    struct llama_context * ctx_llama = llama_init_from_file(params.model_llama.c_str(), lparams);
-
-    // print some info about the processing
-    {
-        fprintf(stderr, "\n");
-
-        if (!whisper_is_multilingual(ctx_wsp)) {
-            if (params.language != "en" || params.translate) {
-                params.language = "en";
-                params.translate = false;
-                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
-            }
-        }
-        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
-                __func__,
-                params.n_threads,
-                params.language.c_str(),
-                params.translate ? "translate" : "transcribe",
-                params.no_timestamps ? 0 : 1);
-
-        fprintf(stderr, "\n");
-    }
-
-
-    // init audio
-
-    audio_async audio(30*1000);
-    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
-        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
-        return 1;
-    }
-
-    audio.resume();
-
-    int n_iter = 0;
-
-    bool is_running  = true;
-    bool force_speak = false;
-
-    float prob0 = 0.0f;
-
-    const std::string chat_symb = ":";
-    const std::string bot_name  = "LLaMA";
-
-    std::vector<float> pcmf32_cur;
-    std::vector<float> pcmf32_prompt;
-
-    const std::string prompt_whisper = ::replace(k_prompt_whisper, "{1}", bot_name);
-
-    // construct the initial prompt for LLaMA inference
-    std::string prompt_llama = params.prompt.empty() ? k_prompt_llama : params.prompt;
-
-    // need to have leading ' '
-    prompt_llama.insert(0, 1, ' ');
-
-    prompt_llama = ::replace(prompt_llama, "{0}", params.person);
-    prompt_llama = ::replace(prompt_llama, "{1}", bot_name);
-
-    {
-        // get time string
-        std::string time_str;
-        {
-            time_t t = time(0);
-            struct tm * now = localtime(&t);
-            char buf[128];
-            strftime(buf, sizeof(buf), "%H:%M", now);
-            time_str = buf;
-        }
-        prompt_llama = ::replace(prompt_llama, "{2}", time_str);
-    }
-
-    {
-        // get year string
-        std::string year_str;
-        {
-            time_t t = time(0);
-            struct tm * now = localtime(&t);
-            char buf[128];
-            strftime(buf, sizeof(buf), "%Y", now);
-            year_str = buf;
-        }
-        prompt_llama = ::replace(prompt_llama, "{3}", year_str);
-    }
-
-    prompt_llama = ::replace(prompt_llama, "{4}", chat_symb);
-
-    // evaluate the initial prompt
-
-    auto embd_inp = ::llama_tokenize(ctx_llama, prompt_llama, true);
-
-    printf("\n");
-    printf("%s : initializing - please wait ...\n", __func__);
-
-    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
-        fprintf(stderr, "%s : failed to eval\n", __func__);
-        return 1;
-    }
-
-    if (params.verbose_prompt) {
-        fprintf(stdout, "\n");
-        fprintf(stdout, "%s", prompt_llama.c_str());
-        fflush(stdout);
-    }
-
-    printf("%s : done! start speaking in the microphone\n", __func__);
-    printf("\n");
-    printf("%s%s", params.person.c_str(), chat_symb.c_str());
-    fflush(stdout);
-
-    // clear audio buffer
-    audio.clear();
-
-    // text inference variables
-    const int voice_id = 2;
-    const int n_keep   = embd_inp.size();
-    const int n_ctx    = llama_n_ctx(ctx_llama);
-
-    int n_past = n_keep;
-    int n_prev = 64; // TODO arg
-
-    std::vector<llama_token> embd;
-
-    // reverse prompts for detecting when it's time to stop speaking
-    std::vector<std::string> antiprompts = {
-        params.person + chat_symb,
-    };
-
-    // main loop
-    while (is_running) {
-        // handle Ctrl + C
-        is_running = sdl_poll_events();
-
-        if (!is_running) {
-            break;
-        }
-
-        // delay
-        std::this_thread::sleep_for(std::chrono::milliseconds(100));
-
-        int64_t t_ms = 0;
-
-        {
-            audio.get(2000, pcmf32_cur);
-
-            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
-                //fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
-
-                audio.get(params.voice_ms, pcmf32_cur);
-
-                std::string text_heard;
-
-                if (!force_speak) {
-                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prompt_whisper, prob0, t_ms));
-                }
-
-                // remove text between brackets using regex
-                {
-                    std::regex re("\\[.*?\\]");
-                    text_heard = std::regex_replace(text_heard, re, "");
-                }
-
-                // remove text between brackets using regex
-                {
-                    std::regex re("\\(.*?\\)");
-                    text_heard = std::regex_replace(text_heard, re, "");
-                }
-
-                // remove all characters, except for letters, numbers, punctuation and ':', '\'', '-', ' '
-                text_heard = std::regex_replace(text_heard, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
-
-                // take first line
-                text_heard = text_heard.substr(0, text_heard.find_first_of('\n'));
-
-                // remove leading and trailing whitespace
-                text_heard = std::regex_replace(text_heard, std::regex("^\\s+"), "");
-                text_heard = std::regex_replace(text_heard, std::regex("\\s+$"), "");
-
-                const std::vector<llama_token> tokens = llama_tokenize(ctx_llama, text_heard.c_str(), false);
-
-                if (text_heard.empty() || tokens.empty() || force_speak) {
-                    //fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
-                    audio.clear();
-
-                    continue;
-                }
-
-                force_speak = false;
-
-                text_heard.insert(0, 1, ' ');
-                text_heard += "\n" + bot_name + chat_symb;
-                fprintf(stdout, "%s%s%s", "\033[1m", text_heard.c_str(), "\033[0m");
-                fflush(stdout);
-
-                embd = ::llama_tokenize(ctx_llama, text_heard, false);
-
-                // text inference
-                bool done = false;
-                std::string text_to_speak;
-                while (true) {
-                    // predict
-                    if (embd.size() > 0) {
-                        if (n_past + (int) embd.size() > n_ctx) {
-                            n_past = n_keep;
-
-                            // insert n_left/2 tokens at the start of embd from last_n_tokens
-                            embd.insert(embd.begin(), embd_inp.begin() + embd_inp.size() - n_prev, embd_inp.end());
-
-                            //printf("\n---\n");
-                            //printf("resetting: '");
-                            //for (int i = 0; i < (int) embd.size(); i++) {
-                            //    printf("%s", llama_token_to_str(ctx_llama, embd[i]));
-                            //}
-                            //printf("'\n");
-                            //printf("\n---\n");
-                        }
-
-                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
-                            fprintf(stderr, "%s : failed to eval\n", __func__);
-                            return 1;
-                        }
-                    }
-
-                    //printf("n_iter = %d, n_past = %d, n_ctx = %d, n_keep = %d, n_prev = %d, embd.size() = %d\n", n_iter, n_past, n_ctx, n_keep, n_prev, (int) embd.size());
-
-                    embd_inp.insert(embd_inp.end(), embd.begin(), embd.end());
-                    n_past += embd.size();
-                    embd.clear();
-
-                    if (done) break;
-
-                    {
-                        // out of user input, sample next token
-                        const float top_k          = 5;
-                        const float top_p          = 0.80f;
-                        const float temp           = 0.30f;
-                        const float repeat_penalty = 1.1764f;
-
-                        const int repeat_last_n    = 256;
-
-                        llama_token id = 0;
-
-                        {
-                            auto logits = llama_get_logits(ctx_llama);
-                            logits[llama_token_eos()] = 0;
-
-                            id = llama_sample_top_p_top_k(ctx_llama,
-                                    embd_inp.data() + std::max(0, n_past - repeat_last_n),
-                                    repeat_last_n, top_k, top_p, temp, repeat_penalty);
-                        }
-
-                        if (id != llama_token_eos()) {
-                            // add it to the context
-                            embd.push_back(id);
-
-                            text_to_speak += llama_token_to_str(ctx_llama, id);
-
-                            printf("%s", llama_token_to_str(ctx_llama, id));
-                        }
-                    }
-
-                    {
-                        std::string last_output;
-                        for (int i = embd_inp.size() - 16; i < (int) embd_inp.size(); i++) {
-                            last_output += llama_token_to_str(ctx_llama, embd_inp[i]);
-                        }
-                        last_output += llama_token_to_str(ctx_llama, embd[0]);
-
-                        for (std::string & antiprompt : antiprompts) {
-                            if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
-                                done = true;
-                                text_to_speak = ::replace(text_to_speak, antiprompt, "");
-                                fflush(stdout);
-                                break;
-                            }
-                        }
-                    }
-
-                    is_running = sdl_poll_events();
-
-                    if (!is_running) {
-                        break;
-                    }
-                }
-
-                text_to_speak = ::replace(text_to_speak, "\"", "");
-                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
-
-                audio.clear();
-
-                ++n_iter;
-            }
-        }
-    }
-
-    audio.pause();
-
-    whisper_print_timings(ctx_wsp);
-    whisper_free(ctx_wsp);
-
-    llama_print_timings(ctx_llama);
-    llama_free(ctx_llama);
-
-    return 0;
-}
--- a/examples/talk.wasm/gpt-2.cpp
+++ b/examples/talk.wasm/gpt-2.cpp
@ -841,7 +841,6 @@ struct gpt2_context * gpt2_init(const char * path_model) {

        if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, "gpt-2.bin");
-            delete ctx;
            return nullptr;
        }

--- a/examples/talk/.gitignore
+++ b/examples/talk/.gitignore
@ -1 +1 @@
-audio.mp3
+eleven-labs.py
--- a/examples/talk/eleven-labs.py
+++ b/examples/talk/eleven-labs.py
@ -1,23 +0,0 @@
-import sys
-import importlib.util
-
-api_key = "" #Write your https://beta.elevenlabs.io api key here
-if not api_key:
-    print("To use elevenlabs you have to register to https://beta.elevenlabs.io and add your elevenlabs api key to examples/talk/eleven-labs.py")
-    sys.exit()
-
-if importlib.util.find_spec("elevenlabs") is None:
-    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
-    sys.exit()
-
-from elevenlabs import ElevenLabs
-eleven = ElevenLabs(api_key)
-
-# Get a Voice object, by name or UUID
-voice = eleven.voices["Arnold"] #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
-
-# Generate the TTS
-audio = voice.generate(str(sys.argv[2:]))
-
-# Save the TTS to a file
-audio.save("audio") 
--- a/examples/talk/speak.sh
+++ b/examples/talk/speak.sh
@ -7,13 +7,9 @@
 # Mac OS: brew install espeak
 # Linux: apt-get install espeak
 #
-#espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
-
-# Mac OS "say" command
-say "$2"
+espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"

 # Eleven Labs
-# To use it, install the elevenlabs module from pip (pip install elevenlabs), register to https://beta.elevenlabs.io to get an api key and paste it in /examples/talk/eleven-labs.py 
 #
 #wd=$(dirname $0)
 #script=$wd/eleven-labs.py
--- a/examples/whisper.swiftui/README.md
+++ b/examples/whisper.swiftui/README.md
@ -1,18 +1,15 @@
 A sample SwiftUI app using [whisper.cpp](https://github.com/ggerganov/whisper.cpp/) to do voice-to-text transcriptions.
 See also: [whisper.objc](https://github.com/ggerganov/whisper.cpp/tree/master/examples/whisper.objc).

-**Usage**:
+To use:

 1. Select a model from the [whisper.cpp repository](https://github.com/ggerganov/whisper.cpp/tree/master/models).[^1]
-2. Add the model to `whisper.swiftui.demo/Resources/models` **via Xcode**.
+2. Add the model to "whisper.swiftui.demo/Resources/models" via Xcode.
 3. Select a sample audio file (for example, [jfk.wav](https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav)).
-4. Add the sample audio file to `whisper.swiftui.demo/Resources/samples` **via Xcode**.
+4. Add the model to "whisper.swiftui.demo/Resources/samples" via Xcode.
 5. Select the "Release" [^2] build configuration under "Run", then deploy and run to your device.

-**Note:** Pay attention to the folder path: `whisper.swiftui.demo/Resources/models` is the appropriate directory to place resources whilst `whisper.swiftui.demo/Models` is related to actual code.
-
 [^1]: I recommend the tiny, base or small models for running on an iOS device.
-
 [^2]: The `Release` build can boost performance of transcription. In this project, it also added `-O3 -DNDEBUG` to `Other C Flags`, but adding flags to app proj is not ideal in real world (applies to all C/C++ files), consider splitting xcodeproj in workspace in your own project.

 ![image](https://user-images.githubusercontent.com/1991296/212539216-0aef65e4-f882-480a-8358-0f816838fd52.png)
--- a/extra/bench-all.sh
+++ b/extra/bench-all.sh
@ -64,10 +64,6 @@ for model in "${models[@]}"; do
        config="$config BLAS"
    fi

-    if [[ $system_info == *"COREML = 1"* ]]; then
-        config="$config COREML"
-    fi
-
    commit=$(git rev-parse --short HEAD)

    printf "| <todo> | <todo> | $config | $model | $n_threads | $load_time | $encode_time | $commit |\n"
--- a/ggml.c
+++ b/ggml.c
--- a/ggml.h
+++ b/ggml.h
@ -177,12 +177,11 @@ extern "C" {
 #include <stddef.h>
 #include <stdbool.h>

-#define GGML_MAX_DIMS          4
-#define GGML_MAX_NODES         4096
-#define GGML_MAX_PARAMS        16
-#define GGML_MAX_CONTEXTS      64
-#define GGML_MAX_OPT           4
-#define GGML_DEFAULT_N_THREADS 4
+#define GGML_MAX_DIMS     4
+#define GGML_MAX_NODES    4096
+#define GGML_MAX_PARAMS   16
+#define GGML_MAX_CONTEXTS 64
+#define GGML_MAX_OPT      4

 #ifdef __ARM_NEON
 // we use the built-in 16-bit float type
@ -199,14 +198,11 @@ struct ggml_object;
 struct ggml_context;

 enum ggml_type {
-    // explicitly numbered values are used in llama.cpp files
-    GGML_TYPE_F32  = 0,
-    GGML_TYPE_F16  = 1,
-    GGML_TYPE_Q4_0 = 2,
-    GGML_TYPE_Q4_1 = 3,
    GGML_TYPE_I8,
    GGML_TYPE_I16,
    GGML_TYPE_I32,
+    GGML_TYPE_F16,
+    GGML_TYPE_F32,
    GGML_TYPE_COUNT,
 };

@ -230,15 +226,12 @@ enum ggml_op {
    GGML_OP_STEP,
    GGML_OP_RELU,
    GGML_OP_GELU,
-    GGML_OP_SILU,
    GGML_OP_NORM, // normalize
-    GGML_OP_RMS_NORM,

    GGML_OP_MUL_MAT,

    GGML_OP_SCALE,
    GGML_OP_CPY,
-    GGML_OP_CONT,
    GGML_OP_RESHAPE,
    GGML_OP_VIEW,
    GGML_OP_PERMUTE,
@ -253,35 +246,19 @@ enum ggml_op {
    GGML_OP_FLASH_ATTN,
    GGML_OP_FLASH_FF,

-    GGML_OP_MAP_UNARY,
-    GGML_OP_MAP_BINARY,
-
    GGML_OP_COUNT,
 };

-
-// ggml object
-struct ggml_object {
-    size_t offs;
-    size_t size;
-
-    struct ggml_object * next;
-
-    char padding[8];
-};
-
-static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
-
 // n-dimensional tensor
 struct ggml_tensor {
    enum ggml_type type;

    int    n_dims;
-    int64_t ne[GGML_MAX_DIMS]; // number of elements
-    size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
-                               // nb[0] = sizeof(type)
-                               // nb[1] = nb[0]   * ne[0] + padding
-                               // nb[i] = nb[i-1] * ne[i-1]
+    int    ne[GGML_MAX_DIMS]; // number of elements
+    size_t nb[GGML_MAX_DIMS]; // stride in bytes:
+                              // nb[0] = sizeof(type)
+                              // nb[1] = nb[0]   * ne[0] + padding
+                              // nb[i] = nb[i-1] * ne[i-1]

    // compute data
    enum ggml_op op;
@ -335,7 +312,6 @@ struct ggml_init_params {
    // memory pool
    size_t mem_size;   // bytes
    void * mem_buffer; // if NULL, memory will be allocated internally
-    bool   no_alloc;   // don't allocate memory for the tensor data
 };

 void    ggml_time_init(void); // call this once at the beginning of the program
@ -347,13 +323,10 @@ int64_t ggml_cycles_per_ms(void);
 void ggml_print_object (const struct ggml_object * obj);
 void ggml_print_objects(const struct ggml_context * ctx);

-int64_t ggml_nelements(const struct ggml_tensor * tensor);
-size_t  ggml_nbytes   (const struct ggml_tensor * tensor);
-
-int    ggml_blck_size (enum ggml_type type);
-size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
-float  ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
+int    ggml_nelements(const struct ggml_tensor * tensor);
+size_t ggml_nbytes   (const struct ggml_tensor * tensor);

+size_t ggml_type_size   (enum ggml_type type);
 size_t ggml_element_size(const struct ggml_tensor * tensor);

 struct ggml_context * ggml_init(struct ggml_init_params params);
@ -367,33 +340,33 @@ struct ggml_tensor * ggml_new_tensor(
        struct ggml_context * ctx,
        enum   ggml_type type,
        int    n_dims,
-        const int64_t *ne);
+        const int *ne);

 struct ggml_tensor * ggml_new_tensor_1d(
        struct ggml_context * ctx,
        enum   ggml_type type,
-        int64_t ne0);
+        int    ne0);

 struct ggml_tensor * ggml_new_tensor_2d(
        struct ggml_context * ctx,
        enum   ggml_type type,
-        int64_t ne0,
-        int64_t ne1);
+        int    ne0,
+        int    ne1);

 struct ggml_tensor * ggml_new_tensor_3d(
        struct ggml_context * ctx,
        enum   ggml_type type,
-        int64_t ne0,
-        int64_t ne1,
-        int64_t ne2);
+        int    ne0,
+        int    ne1,
+        int    ne2);

 struct ggml_tensor * ggml_new_tensor_4d(
        struct ggml_context * ctx,
        enum   ggml_type type,
-        int64_t ne0,
-        int64_t ne1,
-        int64_t ne2,
-        int64_t ne3);
+        int    ne0,
+        int    ne1,
+        int    ne2,
+        int    ne3);

 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
 struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
@ -493,20 +466,12 @@ struct ggml_tensor * ggml_gelu(
        struct ggml_context * ctx,
        struct ggml_tensor  * a);

-struct ggml_tensor * ggml_silu(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a);
-
 // normalize along rows
 // TODO: eps is hardcoded to 1e-5 for now
 struct ggml_tensor * ggml_norm(
        struct ggml_context * ctx,
        struct ggml_tensor  * a);

-struct ggml_tensor * ggml_rms_norm(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a);
-
 // A: m rows, n columns
 // B: p rows, n columns (i.e. we transpose it internally)
 // result is m columns, p rows
@ -531,11 +496,6 @@ struct ggml_tensor * ggml_cpy(
        struct ggml_tensor  * a,
        struct ggml_tensor  * b);

-// make contiguous
-struct ggml_tensor * ggml_cont(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a);
-
 // return view(a), b specifies the new shape
 // TODO: when we start computing gradient, make a copy instead of view
 struct ggml_tensor * ggml_reshape(
@ -548,43 +508,33 @@ struct ggml_tensor * ggml_reshape(
 struct ggml_tensor * ggml_reshape_2d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
-        int64_t               ne0,
-        int64_t               ne1);
+        int                   ne0,
+        int                   ne1);

 // return view(a)
 // TODO: when we start computing gradient, make a copy instead of view
 struct ggml_tensor * ggml_reshape_3d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
-        int64_t               ne0,
-        int64_t               ne1,
-        int64_t               ne2);
+        int                   ne0,
+        int                   ne1,
+        int                   ne2);

 // offset in bytes
 struct ggml_tensor * ggml_view_1d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
-        int64_t               ne0,
+        int                   ne0,
        size_t                offset);

 struct ggml_tensor * ggml_view_2d(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
-        int64_t               ne0,
-        int64_t               ne1,
+        int                   ne0,
+        int                   ne1,
        size_t                nb1, // row stride in bytes
        size_t                offset);

-struct ggml_tensor * ggml_view_3d(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        int64_t               ne0,
-        int64_t               ne1,
-        int64_t               ne2,
-        size_t                nb1, // row   stride in bytes
-        size_t                nb2, // slice stride in bytes
-        size_t                offset);
-
 struct ggml_tensor * ggml_permute(
        struct ggml_context * ctx,
        struct ggml_tensor  * a,
@ -655,21 +605,6 @@ struct ggml_tensor * ggml_flash_ff(
        struct ggml_tensor  * c0,
        struct ggml_tensor  * c1);

-// Mapping operations
-typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
-typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
-
-struct ggml_tensor * ggml_map_unary_f32(
-        struct ggml_context        * ctx,
-        struct ggml_tensor         * a,
-        const  ggml_unary_op_f32_t fun);
-
-struct ggml_tensor * ggml_map_binary_f32(
-        struct ggml_context         * ctx,
-        struct ggml_tensor          * a,
-        struct ggml_tensor          * b,
-        const  ggml_binary_op_f32_t fun);
-
 //
 // automatic differentiation
 //
@ -791,13 +726,6 @@ enum ggml_opt_result ggml_opt(
        struct ggml_opt_params params,
        struct ggml_tensor * f);

-//
-// quantization
-//
-
-size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
-size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
-
 //
 // system info
 //
@ -815,30 +743,6 @@ int ggml_cpu_has_blas(void);
 int ggml_cpu_has_sse3(void);
 int ggml_cpu_has_vsx(void);

-
-//
-// Internal types and functions exposed for tests and benchmarks
-//
-
-#ifdef  __cplusplus
-// restrict not standard in C++
-#define GGML_RESTRICT
-#else
-#define GGML_RESTRICT restrict
-#endif
-typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-typedef void (*quantize_row_q_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-typedef void (*vec_dot_q_t)(const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
-
-typedef struct {
-    dequantize_row_q_t dequantize_row_q;
-    quantize_row_q_t   quantize_row_q;
-    quantize_row_q_t   quantize_row_q_reference;
-    vec_dot_q_t        vec_dot_q;
-} quantize_fns_t;
-
-quantize_fns_t ggml_internal_get_quantize_fn(size_t i);
-
 #ifdef  __cplusplus
 }
 #endif
--- a/models/convert-pt-to-ggml.py
+++ b/models/convert-pt-to-ggml.py
@ -39,7 +39,6 @@ import json
 import code
 import torch
 import numpy as np
-import base64

 #from transformers import GPTJForCausalLM
 #from transformers import GPT2TokenizerFast
@ -225,14 +224,18 @@ with np.load(os.path.join(dir_whisper, "whisper/assets", "mel_filters.npz")) as
 #code.interact(local=locals())

 multilingual = hparams["n_vocab"] == 51865
-tokenizer = os.path.join(dir_whisper, "whisper/assets", multilingual and "multilingual.tiktoken" or "gpt2.tiktoken")
+dir_tokenizer = os.path.join(dir_whisper, "whisper/assets", multilingual and "multilingual" or "gpt2")
+
+#tokenizer = build_tokenizer(dir_whisper, multilingual and "multilingual" or "gpt2")
+#print(tokenizer)
+#print(tokenizer.name_or_path)
+#print(len(tokenizer.additional_special_tokens))

 # output in the same directory as the model
 fname_out = dir_out + "/ggml-model.bin"

-with open(tokenizer, "rb") as f:
-    contents = f.read()
-    tokens = {base64.b64decode(token): int(rank) for token, rank in (line.split() for line in contents.splitlines() if line)}
+with open(dir_tokenizer + "/vocab.json", "r", encoding="utf8") as f:
+    tokens = json.load(f)

 # use 16-bit or 32-bit floats
 use_f16 = True
@ -268,8 +271,9 @@ byte_decoder = {v:k for k, v in byte_encoder.items()}
 fout.write(struct.pack("i", len(tokens)))

 for key in tokens:
-    fout.write(struct.pack("i", len(key)))
-    fout.write(key)
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)

 for name in list_vars.keys():
    data = list_vars[name].squeeze().numpy()
--- a/models/convert-whisper-to-coreml.py
+++ b/models/convert-whisper-to-coreml.py
@ -1,334 +0,0 @@
-import argparse
-import torch
-import torch.nn.functional as F
-import coremltools as ct
-
-from torch import Tensor
-from torch import nn
-from typing import Dict
-from typing import Optional
-from ane_transformers.reference.layer_norm import LayerNormANE as LayerNormANEBase
-from coremltools.models.neural_network.quantization_utils import quantize_weights
-from whisper.model import Whisper, AudioEncoder, TextDecoder, ResidualAttentionBlock, MultiHeadAttention, ModelDimensions
-from whisper import load_model
-
-# Use for changing dim of input in encoder and decoder embeddings
-def linear_to_conv2d_map(state_dict, prefix, local_metadata, strict,
-                         missing_keys, unexpected_keys, error_msgs):
-    """
-    Unsqueeze twice to map nn.Linear weights to nn.Conv2d weights
-    """
-    for k in state_dict:
-        is_attention = all(substr in k for substr in ['attn', '.weight'])
-        is_mlp = any([k.endswith(s) for s in ['mlp.0.weight', 'mlp.2.weight']])
-
-        if (is_attention or is_mlp) and len(state_dict[k].shape) == 2:
-            state_dict[k] = state_dict[k][:, :, None, None]
-
-
-def correct_for_bias_scale_order_inversion(state_dict, prefix, local_metadata,
-                                           strict, missing_keys,
-                                           unexpected_keys, error_msgs):
-    state_dict[prefix + 'bias'] = state_dict[prefix + 'bias'] / state_dict[prefix + 'weight']
-    return state_dict
-
-class LayerNormANE(LayerNormANEBase):
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self._register_load_state_dict_pre_hook(
-            correct_for_bias_scale_order_inversion)
-
-class MultiHeadAttentionANE(MultiHeadAttention):
-    def __init__(self, n_state: int, n_head: int):
-        super().__init__(n_state, n_head)
-
-        setattr(self, 'query', nn.Conv2d(n_state, n_state, kernel_size=1))
-        setattr(self, 'key', nn.Conv2d(n_state, n_state, kernel_size=1, bias=False))
-        setattr(self, 'value', nn.Conv2d(n_state, n_state, kernel_size=1))
-        setattr(self, 'out', nn.Conv2d(n_state, n_state, kernel_size=1))
-
-    def forward(self,
-                x: Tensor,
-                xa: Optional[Tensor] = None,
-                mask: Optional[Tensor] = None,
-                kv_cache: Optional[dict] = None):
-
-        q = self.query(x)
-
-        if kv_cache is None or xa is None or self.key not in kv_cache:
-            # hooks, if installed (i.e. kv_cache is not None), will prepend the cached kv tensors;
-            # otherwise, perform key/value projections for self- or cross-attention as usual.
-            k = self.key(x if xa is None else xa)
-            v = self.value(x if xa is None else xa)
-
-        else:
-            # for cross-attention, calculate keys and values once and reuse in subsequent calls.
-            k = kv_cache[self.key]
-            v = kv_cache[self.value]
-
-        wv, qk = self.qkv_attention_ane(q, k, v, mask)
-
-        return self.out(wv), qk
-
-    def qkv_attention_ane(self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None):
-
-        _, dim, _, seqlen = q.size()
-
-        dim_per_head = dim // self.n_head
-
-        scale = float(dim_per_head)**-0.5
-
-        q = q * scale
-
-        mh_q = q.split(dim_per_head, dim=1)
-        mh_k = k.transpose(1,3).split(dim_per_head, dim=3)
-        mh_v = v.split(dim_per_head, dim=1)
-
-        mh_qk = [
-            torch.einsum('bchq,bkhc->bkhq', [qi, ki])
-            for qi, ki in zip(mh_q, mh_k)
-        ]  # (batch_size, max_seq_length, 1, max_seq_length) * n_heads
-
-        if mask is not None:
-            for head_idx in range(self.n_head):
-                mh_qk[head_idx] = mh_qk[head_idx] + mask[:, :seqlen, :, :seqlen]
-
-        attn_weights = [aw.softmax(dim=1) for aw in mh_qk]  # (batch_size, max_seq_length, 1, max_seq_length) * n_heads
-        attn = [torch.einsum('bkhq,bchk->bchq', wi, vi) for wi, vi in zip(attn_weights, mh_v)]  # (batch_size, dim_per_head, 1, max_seq_length) * n_heads
-        attn = torch.cat(attn, dim=1)  # (batch_size, dim, 1, max_seq_length)
-
-        return attn, torch.cat(mh_qk, dim=1).float().detach()
-
-
-class ResidualAttentionBlockANE(ResidualAttentionBlock):
-    def __init__(self, n_state: int, n_head: int, cross_attention: bool = False):
-        super().__init__(n_state, n_head, cross_attention)
-
-        setattr(self, 'attn', MultiHeadAttentionANE(n_state, n_head))
-        setattr(self, 'attn_ln', LayerNormANE(n_state))
-
-        setattr(self, 'cross_attn', MultiHeadAttentionANE(n_state, n_head) if cross_attention else None)
-        setattr(self, 'cross_attn_ln', LayerNormANE(n_state) if cross_attention else None)
-
-        n_mlp = n_state * 4
-        setattr(self, 'mlp', nn.Sequential(
-            nn.Conv2d(n_state, n_mlp, kernel_size=1),
-            nn.GELU(),
-            nn.Conv2d(n_mlp, n_state, kernel_size=1)
-        ))
-        setattr(self, 'mlp_ln', LayerNormANE(n_state))
-
-
-class AudioEncoderANE(AudioEncoder):
-    def __init__(self, n_mels: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
-        super().__init__(n_mels, n_ctx, n_state, n_head, n_layer)
-
-        setattr(self, 'blocks', nn.ModuleList(
-            [ResidualAttentionBlockANE(n_state, n_head) for _ in range(n_layer)]
-        ))
-        setattr(self, 'ln_post', LayerNormANE(n_state))
-
-    def forward(self, x: Tensor):
-        """
-        x : torch.Tensor, shape = (batch_size, n_mels, n_ctx)
-            the mel spectrogram of the audio
-        """
-        x = F.gelu(self.conv1(x))
-        x = F.gelu(self.conv2(x))
-
-        assert x.shape[1:] == self.positional_embedding.shape[::-1], "incorrect audio shape"
-
-        # Add positional embedding and add dummy dim for ANE
-        x = (x + self.positional_embedding.transpose(0,1)).to(x.dtype).unsqueeze(2)
-
-        for block in self.blocks:
-            x = block(x)
-
-        x = self.ln_post(x)
-
-        # """
-        # TODO:
-        # I think we need to transpose the result here to make it fit whisper.cpp memory order.
-        # However, even doing this, the results are still wrong. Kind of less wrong compared to
-        # not transposing, but still wrong.
-
-        # Also, I don't know why the original OpenAI implementation does not need to transpose
-
-        # transpose to (batch_size, n_ctx, n_state)
-        # x : torch.Tensor, shape = (batch_size, n_state, 1, n_ctx)
-
-        # """
-        # x = x.transpose(1,3)
-
-        return x
-
-class TextDecoderANE(TextDecoder):
-
-    def __init__(self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
-        super().__init__(n_vocab, n_ctx, n_state, n_head, n_layer)
-
-        setattr(self, 'blocks', nn.ModuleList(
-            [ResidualAttentionBlockANE(n_state, n_head, cross_attention=True) for _ in range(n_layer)]
-        ))
-        setattr(self, 'ln', LayerNormANE(n_state))
-
-    def forward(self, x: Tensor, xa: Tensor, kv_cache: Optional[dict] = None):
-        """
-        x : torch.LongTensor, shape = (batch_size, <= n_ctx)
-            the text tokens
-        xa : torch.Tensor, shape = (batch_size, n_mels, n_audio_ctx)
-            the encoded audio features to be attended on
-        """
-        offset = next(iter(kv_cache.values())).shape[3] if kv_cache else 0
-        x = self.token_embedding(x) + self.positional_embedding[offset : offset + x.shape[-1]]
-        x = x.to(xa.dtype)
-
-        # Reformat for ANE
-        mask = self.mask[None, None, :, :].permute(0,3,1,2)
-        x = x.transpose(1,2).unsqueeze(2)
-
-        for block in self.blocks:
-            x = block(x, xa, mask=mask, kv_cache=kv_cache)
-
-        x = self.ln(x)
-
-        # Reformat back from ANE
-        x = x.permute(0,2,3,1).squeeze(0)
-
-        # ANE can only load tensors with dim size of at most 16,384 - whisper uses 51,864 (en) or 51,865 (multi-lang) tokens so we need to compute in chunks
-        if self.token_embedding.weight.shape[0] == 51865:
-            # split in 11 chunks - 4715 each
-            splits = self.token_embedding.weight.split(self.token_embedding.weight.shape[0]//11, dim=0)
-            logits = torch.cat([torch.einsum('bid,jd->bij', x, split) for split in splits]).view(*x.shape[:2], -1)
-        else:
-            # split in 12 chunks - 4322 each
-            assert(self.token_embedding.weight.shape[0] == 51864)
-            splits = self.token_embedding.weight.split(self.token_embedding.weight.shape[0]//12, dim=0)
-            logits = torch.cat([torch.einsum('bid,jd->bij', x, split) for split in splits]).view(*x.shape[:2], -1)
-
-        return logits
-
-class WhisperANE(Whisper):
-    def __init__(self, dims: ModelDimensions):
-        super().__init__(dims)
-
-        setattr(self, 'encoder', AudioEncoderANE(
-            self.dims.n_mels,
-            self.dims.n_audio_ctx,
-            self.dims.n_audio_state,
-            self.dims.n_audio_head,
-            self.dims.n_audio_layer,
-        ))
-        setattr(self, 'decoder', TextDecoderANE(
-            self.dims.n_vocab,
-            self.dims.n_text_ctx,
-            self.dims.n_text_state,
-            self.dims.n_text_head,
-            self.dims.n_text_layer,
-        ))
-
-        self._register_load_state_dict_pre_hook(linear_to_conv2d_map)
-
-    def forward(self, mel: torch.Tensor, tokens: torch.Tensor) -> Dict[str, torch.Tensor]:
-        return self.decoder(tokens, self.encoder(mel))
-
-    def install_kv_cache_hooks(self, cache: Optional[dict] = None):
-        cache = {**cache} if cache is not None else {}
-        hooks = []
-
-        def save_to_cache(module, _, output):
-            if module not in cache or output.shape[3] > self.decoder.positional_embedding.shape[0]:
-                cache[module] = output  # save as-is, for the first token or cross attention
-            else:
-                cache[module] = torch.cat([cache[module], output], dim=3).detach()
-            return cache[module]
-
-        def install_hooks(layer: nn.Module):
-            if isinstance(layer, MultiHeadAttentionANE):
-                hooks.append(layer.key.register_forward_hook(save_to_cache))
-                hooks.append(layer.value.register_forward_hook(save_to_cache))
-
-        self.decoder.apply(install_hooks)
-        return cache, hooks
-
-def convert_encoder(hparams, model, quantize=False):
-    model.eval()
-
-    input_shape = (1, 80, 3000)
-    input_data = torch.randn(input_shape)
-    traced_model = torch.jit.trace(model, input_data)
-
-    model = ct.convert(
-        traced_model,
-        convert_to=None if quantize else "mlprogram", # convert will fail if weights are quantized, not sure why
-        inputs=[ct.TensorType(name="logmel_data", shape=input_shape)],
-        outputs=[ct.TensorType(name="output")],
-        compute_units=ct.ComputeUnit.ALL
-    )
-
-    if quantize:
-        model = quantize_weights(model, nbits=16)
-
-    return model
-
-def convert_decoder(hparams, model, quantize=False):
-    model.eval()
-
-    tokens_shape = (1, 1)
-    audio_shape = (1, hparams.n_audio_state, 1, 1500)
-
-    audio_data = torch.randn(audio_shape)
-    token_data = torch.randint(50257, tokens_shape).long()
-    traced_model = torch.jit.trace(model, (token_data, audio_data))
-
-    model = ct.convert(
-        traced_model,
-        convert_to=None if quantize else "mlprogram", # convert will fail if weights are quantized, not sure why
-        inputs=[
-            ct.TensorType(name="token_data", shape=tokens_shape, dtype=int),
-            ct.TensorType(name="audio_data", shape=audio_shape)
-        ]
-    )
-
-    if quantize:
-        model = quantize_weights(model, nbits=16)
-
-    return model
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--model", type=str, help="model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large)", required=True)
-    parser.add_argument("--encoder-only", type=bool, help="only convert encoder", default=False)
-    parser.add_argument("--quantize",     type=bool, help="quantize weights to F16", default=False)
-    parser.add_argument("--optimize-ane", type=bool, help="optimize for ANE execution (currently broken)", default=False)
-    args = parser.parse_args()
-
-    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large"]:
-        raise ValueError("Invalid model name")
-
-    whisper = load_model(args.model).cpu()
-    hparams = whisper.dims
-    print(hparams)
-
-    if args.optimize_ane:
-        whisperANE = WhisperANE(hparams).eval()
-        whisperANE.load_state_dict(whisper.state_dict())
-
-        encoder = whisperANE.encoder
-        decoder = whisperANE.decoder
-    else:
-        encoder = whisper.encoder
-        decoder = whisper.decoder
-
-    # Convert encoder
-    encoder = convert_encoder(hparams, encoder, quantize=args.quantize)
-    encoder.save(f"models/coreml-encoder-{args.model}.mlpackage")
-
-    if args.encoder_only is False:
-        # Convert decoder
-        decoder = convert_decoder(hparams, decoder, quantize=args.quantize)
-        decoder.save(f"models/coreml-decoder-{args.model}.mlpackage")
-
-    print("done converting")
--- a/models/download-ggml-model.sh
+++ b/models/download-ggml-model.sh
@ -12,7 +12,7 @@ pfx="resolve/main/ggml"
 # get the path of this script
 function get_script_path() {
    if [ -x "$(command -v realpath)" ]; then
-        echo "$(dirname "$(realpath "$0")")"
+        echo "$(dirname $(realpath $0))"
    else
        local ret="$(cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P)"
        echo "$ret"
--- a/models/generate-coreml-interface.sh
+++ b/models/generate-coreml-interface.sh
@ -1,29 +0,0 @@
-#!/bin/bash
-#
-# This generates:
-#   - coreml/whisper-encoder-impl.h and coreml/whisper-encoder-impl.m
-#   - coreml/whisper-decoder-impl.h and coreml/whisper-decoder-impl.m
-#
-
-wd=$(dirname "$0")
-cd "$wd/../"
-
-python3 models/convert-whisper-to-coreml.py --model tiny.en
-
-mv -v models/coreml-encoder-tiny.en.mlpackage models/whisper-encoder-impl.mlpackage
-xcrun coremlc generate models/whisper-encoder-impl.mlpackage coreml/
-mv coreml/whisper_encoder_impl.h coreml/whisper-encoder-impl.h
-mv coreml/whisper_encoder_impl.m coreml/whisper-encoder-impl.m
-sed -i '' 's/whisper_encoder_impl\.h/whisper-encoder-impl.h/g' coreml/whisper-encoder-impl.m
-sed -i '' 's/whisper_encoder_impl\.m/whisper-encoder-impl.m/g' coreml/whisper-encoder-impl.m
-sed -i '' 's/whisper_encoder_impl\.h/whisper-encoder-impl.h/g' coreml/whisper-encoder-impl.h
-
-mv -v models/coreml-decoder-tiny.en.mlpackage models/whisper-decoder-impl.mlpackage
-xcrun coremlc generate models/whisper-decoder-impl.mlpackage coreml/
-mv coreml/whisper_decoder_impl.h coreml/whisper-decoder-impl.h
-mv coreml/whisper_decoder_impl.m coreml/whisper-decoder-impl.m
-sed -i '' 's/whisper_decoder_impl\.h/whisper-decoder-impl.h/g' coreml/whisper-decoder-impl.m
-sed -i '' 's/whisper_decoder_impl\.m/whisper-decoder-impl.m/g' coreml/whisper-decoder-impl.m
-sed -i '' 's/whisper_decoder_impl\.h/whisper-decoder-impl.h/g' coreml/whisper-decoder-impl.h
-
-rm -rfv models/whisper-encoder-impl.mlpackage models/whisper-decoder-impl.mlpackage
--- a/models/generate-coreml-model.sh
+++ b/models/generate-coreml-model.sh
@ -1,25 +0,0 @@
-#!/bin/bash
-
-# Usage: ./generate-coreml-model.sh <model-name>
-if [ $# -eq 0 ]
-  then
-    echo "No model name supplied"
-    echo "Usage: ./generate-coreml-model.sh <model-name>"
-    exit 1
-fi
-
-mname="$1"
-
-wd=$(dirname "$0")
-cd "$wd/../"
-
-python3 models/convert-whisper-to-coreml.py --model $mname --encoder-only True
-
-xcrun coremlc compile models/coreml-encoder-${mname}.mlpackage models/
-rm -rf models/ggml-${mname}-encoder.mlmodelc
-mv -v models/coreml-encoder-${mname}.mlmodelc models/ggml-${mname}-encoder.mlmodelc
-
-# TODO: decoder (sometime in the future maybe)
-#xcrun coremlc compile models/whisper-decoder-${mname}.mlpackage models/
-#rm -rf models/ggml-${mname}-decoder.mlmodelc
-#mv -v models/coreml_decoder_${mname}.mlmodelc models/ggml-${mname}-decoder.mlmodelc
--- a/whisper.cpp
+++ b/whisper.cpp
@ -221,14 +221,14 @@ static const std::map<std::string, std::pair<int, std::string>> g_lang = {
    { "su",  { 98,  "sundanese",      } },
 };

-static const size_t MB = 1ull*1024*1024;
+static const size_t MB = 1024*1024;

 static const std::map<e_model, size_t> MEM_REQ_SCRATCH0 = {
-    { MODEL_TINY,     14ull*MB },
-    { MODEL_BASE,     18ull*MB },
-    { MODEL_SMALL,    28ull*MB },
-    { MODEL_MEDIUM,   36ull*MB },
-    { MODEL_LARGE,    44ull*MB },
+    { MODEL_TINY,     12ull*MB },
+    { MODEL_BASE,     15ull*MB },
+    { MODEL_SMALL,    23ull*MB },
+    { MODEL_MEDIUM,   31ull*MB },
+    { MODEL_LARGE,    38ull*MB },
 };

 static const std::map<e_model, size_t> MEM_REQ_SCRATCH1 = {
@ -297,7 +297,6 @@ static const std::map<e_model, size_t> MEM_REQ_DECODE = {

 struct whisper_mel {
    int n_len;
-    int n_len_org;
    int n_mel;

    std::vector<float> data;
@ -590,7 +589,6 @@ struct whisper_state {

    int lang_id = 0; // english by default

-    std::string path_model; // populated by whisper_init_from_file()
 #ifdef WHISPER_USE_COREML
    whisper_coreml_context * ctx_coreml;
 #endif
@ -663,11 +661,9 @@ static bool kv_cache_init(
                                 int   n_ctx) {
    cache.buf.resize(mem_bytes);

-    struct ggml_init_params params = {
-        /*.mem_size   =*/ cache.buf.size(),
-        /*.mem_buffer =*/ cache.buf.data(),
-        /*.no_alloc   =*/ false,
-    };
+    struct ggml_init_params params;
+    params.mem_size   = cache.buf.size();
+    params.mem_buffer = cache.buf.data();

    cache.ctx = ggml_init(params);

@ -699,11 +695,9 @@ static bool kv_cache_reinit(struct whisper_kv_cache & cache) {

    WHISPER_ASSERT(cache.buf.size() >= 2*n_elements*ggml_type_size(wtype));

-    struct ggml_init_params params = {
-        /*.mem_size   =*/ cache.buf.size(),
-        /*.mem_buffer =*/ cache.buf.data(),
-        /*.no_alloc   =*/ false,
-    };
+    struct ggml_init_params params;
+    params.mem_size   = cache.buf.size();
+    params.mem_buffer = cache.buf.data();

    cache.ctx = ggml_init(params);

@ -1041,11 +1035,9 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con

    // create the ggml context
    {
-        struct ggml_init_params params = {
-            /*.mem_size   =*/ wctx.model.buf->size(),
-            /*.mem_buffer =*/ wctx.model.buf->data(),
-            /*.no_alloc   =*/ false,
-        };
+        struct ggml_init_params params;
+        params.mem_size   = wctx.model.buf->size();
+        params.mem_buffer = wctx.model.buf->data();

        model.ctx = ggml_init(params);
        if (!model.ctx) {
@ -1294,7 +1286,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con

            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1] || tensor->ne[2] != ne[2]) {
                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d, %d], expected [%d, %d, %d]\n",
-                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], ne[0], ne[1], ne[2]);
+                        __func__, name.data(), tensor->ne[0], tensor->ne[1], tensor->ne[2], ne[0], ne[1], ne[2]);
                return false;
            }

@ -1359,11 +1351,9 @@ static bool whisper_encode_internal(
    const int n_mels = hparams.n_mels;
    assert(mel_inp.n_mel == n_mels);

-    struct ggml_init_params params = {
-        /*.mem_size   =*/ wstate.buf_compute.size(),
-        /*.mem_buffer =*/ wstate.buf_compute.data(),
-        /*.no_alloc   =*/ false,
-    };
+    struct ggml_init_params params;
+    params.mem_size   = wstate.buf_compute.size();
+    params.mem_buffer = wstate.buf_compute.data();

    struct ggml_context * ctx0 = ggml_init(params);

@ -1519,7 +1509,8 @@ static bool whisper_encode_internal(
                                Vcur,
                                n_state/n_head, n_head, n_ctx),
                            1, 2, 0, 3),
-                        ggml_new_tensor_3d(ctx0, wctx.wtype, n_ctx, n_state/n_head, n_head));
+                        ggml_new_tensor_3d(ctx0, wctx.wtype, n_ctx, n_state/n_head, n_head)
+                        );

            struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, false);
 #else
@ -1750,12 +1741,10 @@ static bool whisper_encode_internal(

            wstate.use_buf(ctx0, -1);

-            Vcross = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcross, n_state, n_ctx));
-
-            struct ggml_tensor * k = ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx));
-            struct ggml_tensor * v = ggml_view_2d(ctx0, wstate.kv_cross.v, n_ctx, n_state,
-                    (   n_ctx)*ggml_element_size(wstate.kv_cross.v),
-                    (il*n_ctx)*ggml_element_size(wstate.kv_cross.v)*n_state);
+            //struct ggml_tensor * k = ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*hparams.n_audio_ctx + iter*n_ctx));
+            //struct ggml_tensor * v = ggml_view_1d(ctx0, wstate.kv_cross.v, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.v)*n_state)*(il*hparams.n_audio_ctx + iter*n_ctx));
+            struct ggml_tensor* k = ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx));
+            struct ggml_tensor* v = ggml_view_1d(ctx0, wstate.kv_cross.v, n_state*n_ctx, (ggml_element_size(wstate.kv_cross.v)*n_state)*(il*n_ctx));

            ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcross, k));
            ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcross, v));
@ -1769,10 +1758,10 @@ static bool whisper_encode_internal(

    //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
    //        ggml_used_mem(ctx0)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
+    //        wctx.get_buf_max_mem(0)/1024.0/1024.0,
+    //        wctx.get_buf_max_mem(1)/1024.0/1024.0,
+    //        wctx.get_buf_max_mem(2)/1024.0/1024.0,
+    //        wctx.get_buf_max_mem(3)/1024.0/1024.0);

    ggml_free(ctx0);

@ -1823,11 +1812,9 @@ static bool whisper_decode_internal(

    //WHISPER_PRINT_DEBUG("%s: n_past = %d, N = %d, M = %d, n_ctx = %d\n", __func__, n_past, N, M, n_ctx);

-    struct ggml_init_params params = {
-        /*.mem_size   =*/ wstate.buf_compute.size(),
-        /*.mem_buffer =*/ wstate.buf_compute.data(),
-        /*.no_alloc   =*/ false,
-    };
+    struct ggml_init_params params;
+    params.mem_size   = wstate.buf_compute.size();
+    params.mem_buffer = wstate.buf_compute.data();

    struct ggml_context * ctx0 = ggml_init(params);

@ -1871,6 +1858,8 @@ static bool whisper_decode_internal(

        // self-attention
        {
+            wstate.use_buf(ctx0, 1);
+
            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0,
                    layer.attn_q_w,
                    cur);
@ -1890,24 +1879,20 @@ static bool whisper_decode_internal(

            Kcur = ggml_scale(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));

+            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0,
+                    layer.attn_v_w,
+                    cur);
+
+            Vcur = ggml_add(ctx0,
+                    ggml_repeat(ctx0,
+                        layer.attn_v_b,
+                        Vcur),
+                    Vcur);
+
            // store key and value to memory
            {
-                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0,
-                        layer.attn_v_w,
-                        cur);
-
-                Vcur = ggml_add(ctx0,
-                        ggml_repeat(ctx0,
-                            layer.attn_v_b,
-                            Vcur),
-                        Vcur);
-
-                Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_state, N));
-
                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_state, (ggml_element_size(kv_self.k)*n_state)*(il*n_ctx + n_past));
-                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_state,
-                        (   n_ctx)*ggml_element_size(kv_self.v),
-                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_state + n_past*ggml_element_size(kv_self.v));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, kv_self.v, N*n_state, (ggml_element_size(kv_self.v)*n_state)*(il*n_ctx + n_past));

                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
@ -1936,6 +1921,8 @@ static bool whisper_decode_internal(
            // K * Q
            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);

+            wstate.use_buf(ctx0, 0);
+
            //struct ggml_tensor * KQ_scaled =
            //    ggml_scale(ctx0,
            //            KQ,
@ -1944,16 +1931,22 @@ static bool whisper_decode_internal(

            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ, n_past);

+            wstate.use_buf(ctx0, 1);
+
            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);

-            struct ggml_tensor * V =
-                ggml_view_3d(ctx0, kv_self.v,
-                        n_past + N, n_state/n_head, n_head,
-                        n_ctx*ggml_element_size(kv_self.v),
-                        n_ctx*ggml_element_size(kv_self.v)*n_state/n_head,
-                        il*n_ctx*ggml_element_size(kv_self.v)*n_state);
+            wstate.use_buf(ctx0, 0);

-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            struct ggml_tensor * V_trans =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_state, il*n_ctx*ggml_element_size(kv_self.v)*n_state),
+                            n_state/n_head, n_head, n_past + N),
+                        1, 2, 0, 3);
+
+            wstate.use_buf(ctx0, 1);
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);

            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);

@ -1988,6 +1981,8 @@ static bool whisper_decode_internal(

            cur = ggml_norm(ctx0, inpCA); // note: we use inpCA here

+            wstate.use_buf(ctx0, 1);
+
            // cur = ln_0_w*cur + ln_0_b
            cur = ggml_add(ctx0,
                    ggml_mul(ctx0,
@ -1998,6 +1993,8 @@ static bool whisper_decode_internal(

        // cross-attention
        {
+            wstate.use_buf(ctx0, 0);
+
            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0,
                    layer.cross_attn_q_w,
                    cur);
@ -2016,25 +2013,17 @@ static bool whisper_decode_internal(
                        ggml_view_1d(ctx0, wstate.kv_cross.k, M*n_state, il*M*ggml_element_size(wstate.kv_cross.k)*n_state),
                        n_state/n_head, n_head, M);

-            //struct ggml_tensor * Vcross =
-            //    ggml_reshape_3d(ctx0,
-            //            ggml_view_1d(ctx0, wstate.kv_cross.v, M*n_state, il*M*ggml_element_size(wstate.kv_cross.v)*n_state),
-            //            n_state/n_head, n_head, M);
+            struct ggml_tensor * Vcross =
+                ggml_reshape_3d(ctx0,
+                        ggml_view_1d(ctx0, wstate.kv_cross.v, M*n_state, il*M*ggml_element_size(wstate.kv_cross.v)*n_state),
+                        n_state/n_head, n_head, M);

-            //struct ggml_tensor * V_trans =
-            //    ggml_cpy(ctx0,
-            //            ggml_permute(ctx0, Vcross, 1, 2, 0, 3),
-            //            ggml_new_tensor_3d(ctx0, Vcross->type, M, n_state/n_head, n_head));
-
-            struct ggml_tensor * V =
-                ggml_view_3d(ctx0, wstate.kv_cross.v,
-                        M, n_state/n_head, n_head,
-                        M*ggml_element_size(wstate.kv_cross.v),
-                        M*ggml_element_size(wstate.kv_cross.v)*n_state/n_head,
-                        il*M*ggml_element_size(wstate.kv_cross.v)*n_state);
+            struct ggml_tensor * V_trans = ggml_permute(ctx0, Vcross, 1, 2, 0, 3);

            // ------

+            wstate.use_buf(ctx0, 1);
+
            struct ggml_tensor * Q =
                ggml_permute(ctx0,
                        ggml_cpy(ctx0,
@ -2044,6 +2033,8 @@ static bool whisper_decode_internal(

            struct ggml_tensor * K = ggml_permute(ctx0, Kcross, 0, 2, 1, 3);

+            wstate.use_buf(ctx0, 0);
+
            // K * Q
            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);

@ -2056,9 +2047,15 @@ static bool whisper_decode_internal(
            // no masking for cross-attention
            //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);

+            wstate.use_buf(ctx0, 1);
+
            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ);

-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            wstate.use_buf(ctx0, 0);
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            wstate.use_buf(ctx0, 1);

            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);

@ -2190,10 +2187,10 @@ static bool whisper_decode_internal(
    if (N > 1) {
        //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
        //        ggml_used_mem(ctx0)/1024.0/1024.0,
-        //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
-        //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
-        //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
-        //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
+        //        wctx.get_buf_max_mem(0)/1024.0/1024.0,
+        //        wctx.get_buf_max_mem(1)/1024.0/1024.0,
+        //        wctx.get_buf_max_mem(2)/1024.0/1024.0,
+        //        wctx.get_buf_max_mem(3)/1024.0/1024.0);
    }

    ggml_free(ctx0);
@ -2301,72 +2298,6 @@ static void fft(const std::vector<float> & in, std::vector<float> & out) {
    }
 }

-static void log_mel_spectrogram_worker_thread(int ith, const std::vector<float> &hann, const float *samples,
-                                              int n_samples, int fft_size, int fft_step, int n_threads,
-                                              const whisper_filters &filters, bool speed_up, whisper_mel &mel) {
-    std::vector<float> fft_in(fft_size, 0.0);
-    std::vector<float> fft_out(2 * fft_size);
-    int n_fft = 1 + (speed_up ? fft_size / 4 : fft_size / 2);
-
-    for (int i = ith; i < mel.n_len; i += n_threads) {
-        const int offset = i * fft_step;
-
-        // apply Hanning window
-        for (int j = 0; j < fft_size; j++) {
-            if (offset + j < n_samples) {
-                fft_in[j] = hann[j] * samples[offset + j];
-            } else {
-                fft_in[j] = 0.0;
-            }
-        }
-
-        // FFT -> mag^2
-        fft(fft_in, fft_out);
-
-        for (int j = 0; j < fft_size; j++) {
-            fft_out[j] = (fft_out[2 * j + 0] * fft_out[2 * j + 0] + fft_out[2 * j + 1] * fft_out[2 * j + 1]);
-        }
-        for (int j = 1; j < fft_size / 2; j++) {
-            fft_out[j] += fft_out[fft_size - j];
-        }
-
-        if (speed_up) {
-            // scale down in the frequency domain results in a speed up in the time domain
-            for (int j = 0; j < n_fft; j++) {
-                fft_out[j] = 0.5 * (fft_out[2 * j] + fft_out[2 * j + 1]);
-            }
-        }
-
-        // mel spectrogram
-        for (int j = 0; j < mel.n_mel; j++) {
-            double sum = 0.0;
-
-            // unroll loop (suggested by GH user @lunixbochs)
-            int k = 0;
-            for (k = 0; k < n_fft - 3; k += 4) {
-                sum +=
-                    fft_out[k + 0] * filters.data[j*n_fft + k + 0] +
-                    fft_out[k + 1] * filters.data[j*n_fft + k + 1] +
-                    fft_out[k + 2] * filters.data[j*n_fft + k + 2] +
-                    fft_out[k + 3] * filters.data[j*n_fft + k + 3];
-            }
-
-            // handle n_fft remainder
-            for (; k < n_fft; k++) {
-                sum += fft_out[k] * filters.data[j * n_fft + k];
-            }
-
-            if (sum < 1e-10) {
-                sum = 1e-10;
-            }
-
-            sum = log10(sum);
-
-            mel.data[j * mel.n_len + i] = sum;
-        }
-    }
-}
-
 // ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L92-L124
 static bool log_mel_spectrogram(
          whisper_state & wstate,
@ -2389,48 +2320,85 @@ static bool log_mel_spectrogram(
        hann[i] = 0.5*(1.0 - cos((2.0*M_PI*i)/(fft_size)));
    }

-    mel.n_mel     = n_mel;
-    mel.n_len     = n_samples/fft_step;
-    mel.n_len_org = mel.n_len;
-
-    std::vector<float> samples_padded;
-
-    // pad audio with at least one extra chunk of zeros
-    {
-        const int pad = (100*WHISPER_CHUNK_SIZE)/2;
-
-        if (mel.n_len % pad != 0) {
-            mel.n_len = (mel.n_len/pad + 1)*pad;
-        }
-        mel.n_len += pad;
-
-        samples_padded.resize(mel.n_len*fft_step);
-        memcpy(samples_padded.data(), samples, n_samples*sizeof(float));
-        memset(samples_padded.data() + n_samples, 0, (mel.n_len*fft_step - n_samples)*sizeof(float));
-
-        samples = samples_padded.data();
-    }
-
+    mel.n_mel = n_mel;
+    mel.n_len = (n_samples)/fft_step;
    mel.data.resize(mel.n_mel*mel.n_len);

+    const int n_fft = 1 + (speed_up ? fft_size/4 : fft_size/2);
+
    //printf("%s: n_samples = %d, n_len = %d\n", __func__, n_samples, mel.n_len);
    //printf("%s: recording length: %f s\n", __func__, (float) n_samples/sample_rate);

-    {
-        std::vector<std::thread> workers(n_threads - 1);
-        for (int iw = 0; iw < n_threads - 1; ++iw) {
-            workers[iw] = std::thread(
-                    log_mel_spectrogram_worker_thread, iw + 1, std::cref(hann), samples,
-                    n_samples, fft_size, fft_step, n_threads,
-                    std::cref(filters), speed_up, std::ref(mel));
-        }
+    std::vector<std::thread> workers(n_threads);
+    for (int iw = 0; iw < n_threads; ++iw) {
+        workers[iw] = std::thread([&](int ith) {
+            std::vector<float> fft_in;
+            fft_in.resize(fft_size);
+            for (int i = 0; i < fft_size; i++) {
+                fft_in[i] = 0.0;
+            }

-        // main thread
-        log_mel_spectrogram_worker_thread(0, hann, samples, n_samples, fft_size, fft_step, n_threads, filters, speed_up, mel);
+            std::vector<float> fft_out;
+            fft_out.resize(2*fft_size);

-        for (int iw = 0; iw < n_threads - 1; ++iw) {
-            workers[iw].join();
-        }
+            for (int i = ith; i < mel.n_len; i += n_threads) {
+                const int offset = i*fft_step;
+
+                // apply Hanning window
+                for (int j = 0; j < fft_size; j++) {
+                    if (offset + j < n_samples) {
+                        fft_in[j] = hann[j]*samples[offset + j];
+                    } else {
+                        fft_in[j] = 0.0;
+                    }
+                }
+
+                // FFT -> mag^2
+                fft(fft_in, fft_out);
+
+                for (int j = 0; j < fft_size; j++) {
+                    fft_out[j] = (fft_out[2*j + 0]*fft_out[2*j + 0] + fft_out[2*j + 1]*fft_out[2*j + 1]);
+                }
+                for (int j = 1; j < fft_size/2; j++) {
+                    //if (i == 0) {
+                    //    printf("%d: %f %f\n", j, fft_out[j], fft_out[fft_size - j]);
+                    //}
+                    fft_out[j] += fft_out[fft_size - j];
+                }
+                if (i == 0) {
+                    //for (int j = 0; j < fft_size; j++) {
+                    //    printf("%d: %e\n", j, fft_out[j]);
+                    //}
+                }
+
+                if (speed_up) {
+                    // scale down in the frequency domain results in a speed up in the time domain
+                    for (int j = 0; j < n_fft; j++) {
+                        fft_out[j] = 0.5*(fft_out[2*j] + fft_out[2*j + 1]);
+                    }
+                }
+
+                // mel spectrogram
+                for (int j = 0; j < mel.n_mel; j++) {
+                    double sum = 0.0;
+
+                    for (int k = 0; k < n_fft; k++) {
+                        sum += fft_out[k]*filters.data[j*n_fft + k];
+                    }
+                    if (sum < 1e-10) {
+                        sum = 1e-10;
+                    }
+
+                    sum = log10(sum);
+
+                    mel.data[j*mel.n_len + i] = sum;
+                }
+            }
+        }, iw);
+    }
+
+    for (int iw = 0; iw < n_threads; ++iw) {
+        workers[iw].join();
    }

    // clamping and normalization
@ -2454,8 +2422,6 @@ static bool log_mel_spectrogram(

    wstate.t_mel_us += ggml_time_us() - t_start_us;

-    //printf("mel.n_len() = %d, divided by 1500: %f, n_samples / fft_step: %d\n", mel.n_len, mel.n_len / 1500.0, n_samples / fft_step);
-
    return true;
 }

@ -2497,20 +2463,25 @@ static std::vector<whisper_vocab::id> tokenize(const whisper_vocab & vocab, cons
        int n = word.size();
        while (i < n) {
            int j = n;
-            bool found = false;
            while (j > i) {
-                auto sub = word.substr(i, j-i);
-                auto it = vocab.token_to_id.find(sub);
+                auto it = vocab.token_to_id.find(word.substr(i, j-i));
                if (it != vocab.token_to_id.end()) {
                    tokens.push_back(it->second);
                    i = j;
-                    found = true;
                    break;
                }
                --j;
            }
-            if (!found) {
-                fprintf(stderr, "unknown token \n");
+            if (i == n) {
+                break;
+            }
+            if (j == i) {
+                auto sub = word.substr(i, 1);
+                if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
+                    tokens.push_back(vocab.token_to_id.at(sub));
+                } else {
+                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
+                }
                ++i;
            }
        }
@ -2524,14 +2495,14 @@ static std::vector<whisper_vocab::id> tokenize(const whisper_vocab & vocab, cons
 //

 #ifdef WHISPER_USE_COREML
-// replace .bin with -encoder.mlmodelc
-static std::string whisper_get_coreml_path_encoder(std::string path_bin) {
+// replace .bin with .mlmodelc
+static std::string whisper_get_coreml_path(std::string path_bin) {
    auto pos = path_bin.rfind('.');
    if (pos != std::string::npos) {
        path_bin = path_bin.substr(0, pos);
    }

-    path_bin += "-encoder.mlmodelc";
+    path_bin += ".mlmodelc";

    return path_bin;
 }
@ -2544,7 +2515,6 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {

    if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_SELF.at(ctx->model.type), state->decoders[0].kv_self, ctx->wtype, ctx->model.hparams.n_text_ctx)) {
        fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
-        delete state;
        return nullptr;
    }

@ -2555,7 +2525,6 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {

    if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_CROSS.at(ctx->model.type), state->kv_cross, ctx->wtype, ctx->model.hparams.n_audio_ctx)) {
        fprintf(stderr, "%s: kv_cache_init() failed for cross-attention cache\n", __func__);
-        delete state;
        return nullptr;
    }

@ -2565,18 +2534,18 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
    }

 #ifdef WHISPER_USE_COREML
-    const auto path_coreml = whisper_get_coreml_path_encoder(ctx->path_model);
+        const auto path_coreml = whisper_get_coreml_path(ctx->path_model);

-    fprintf(stderr, "%s: loading Core ML model from '%s'\n", __func__, path_coreml.c_str());
-    fprintf(stderr, "%s: first run on a device may take a while ...\n", __func__);
+        fprintf(stderr, "%s: loading Core ML model from '%s'\n", __func__, path_coreml.c_str());
+        fprintf(stderr, "%s: first run on a device may take a while ...\n", __func__);

-    state->ctx_coreml = whisper_coreml_init(path_coreml.c_str());
-    if (!state->ctx_coreml) {
-        fprintf(stderr, "%s: failed to load Core ML model from '%s'\n", __func__, path_coreml.c_str());
-        return nullptr;
-    }
+        state->ctx_coreml = whisper_coreml_init(path_coreml.c_str());
+        if (!state->ctx_coreml) {
+            fprintf(stderr, "%s: failed to load Core ML model from '%s'\n", __func__, path_coreml.c_str());
+            return nullptr;
+        }

-    fprintf(stderr, "%s: Core ML model loaded\n", __func__);
+        fprintf(stderr, "%s: Core ML model loaded\n", __func__);
 #endif

    state->logits.reserve(ctx->vocab.n_vocab * ctx->model.hparams.n_text_ctx);
@ -2746,11 +2715,6 @@ void whisper_free_state(struct whisper_state * state)
            kv_cache_free(state->decoders[i].kv_self);
        }

-#ifdef WHISPER_USE_COREML
-        whisper_coreml_free(state->ctx_coreml);
-        state->ctx_coreml = nullptr;
-#endif
-
        delete state;
    }
 }
@ -2766,6 +2730,10 @@ void whisper_free(struct whisper_context * ctx) {

        whisper_free_state(ctx->state);

+#ifdef WHISPER_USE_COREML
+        whisper_coreml_free(ctx->state->ctx_coreml);
+        ctx->state->ctx_coreml = nullptr;
+#endif
        delete ctx;
    }
 }
@ -2809,9 +2777,8 @@ int whisper_set_mel_with_state(
        return -1;
    }

-    state->mel.n_len     = n_len;
-    state->mel.n_len_org = n_len;
-    state->mel.n_mel     = n_mel;
+    state->mel.n_len = n_len;
+    state->mel.n_mel = n_mel;

    state->mel.data.resize(n_len*n_mel);
    memcpy(state->mel.data.data(), data, n_len*n_mel*sizeof(float));
@ -2937,8 +2904,8 @@ int whisper_lang_auto_detect_with_state(
        return -1;
    }

-    if (seek >= state->mel.n_len_org) {
-        fprintf(stderr, "%s: offset %dms is past the end of the audio (%dms)\n", __func__, offset_ms, state->mel.n_len_org*10);
+    if (seek >= state->mel.n_len) {
+        fprintf(stderr, "%s: offset %dms is past the end of the audio (%dms)\n", __func__, offset_ms, state->mel.n_len*10);
        return -2;
    }

@ -3073,11 +3040,11 @@ const char *whisper_model_type_readable(struct whisper_context * ctx) {
 }

 int whisper_n_len_from_state(struct whisper_state * state) {
-    return state->mel.n_len_org;
+    return state->mel.n_len;
 }

 int whisper_n_len(struct whisper_context * ctx) {
-    return ctx->state->mel.n_len_org;
+    return ctx->state->mel.n_len;
 }

 int whisper_n_vocab(struct whisper_context * ctx) {
@ -3173,14 +3140,6 @@ void whisper_reset_timings(struct whisper_context * ctx) {
    }
 }

-static int whisper_has_coreml(void) {
-#ifdef WHISPER_USE_COREML
-    return 1;
-#else
-    return 0;
-#endif
-}
-
 const char * whisper_print_system_info(void) {
    static std::string s;

@ -3197,7 +3156,6 @@ const char * whisper_print_system_info(void) {
    s += "BLAS = "      + std::to_string(ggml_cpu_has_blas())      + " | ";
    s += "SSE3 = "      + std::to_string(ggml_cpu_has_sse3())      + " | ";
    s += "VSX = "       + std::to_string(ggml_cpu_has_vsx())       + " | ";
-    s += "COREML = "    + std::to_string(whisper_has_coreml())     + " | ";

    return s.c_str();
 }
@ -3231,7 +3189,6 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
        /*.speed_up         =*/ false,
        /*.audio_ctx        =*/ 0,

-        /*.initial_prompt   =*/ nullptr,
        /*.prompt_tokens    =*/ nullptr,
        /*.prompt_n_tokens  =*/ 0,

@ -3244,7 +3201,7 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
        /*.max_initial_ts   =*/  1.0f,
        /*.length_penalty   =*/ -1.0f,

-        /*.temperature_inc  =*/  0.4f,
+        /*.temperature_inc  =*/  0.0f, // TODO: temporary disabled until improve performance
        /*.entropy_thold    =*/  2.4f,
        /*.logprob_thold    =*/ -1.0f,
        /*.no_speech_thold  =*/  0.6f,
@ -3262,9 +3219,6 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
        /*.new_segment_callback           =*/ nullptr,
        /*.new_segment_callback_user_data =*/ nullptr,

-        /*.progress_callback           =*/ nullptr,
-        /*.progress_callback_user_data =*/ nullptr,
-
        /*.encoder_begin_callback           =*/ nullptr,
        /*.encoder_begin_callback_user_data =*/ nullptr,

@ -3276,13 +3230,13 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
        case WHISPER_SAMPLING_GREEDY:
            {
                result.greedy = {
-                    /*.best_of   =*/ 2, // TODO: increase to 5 when we speed-up batch decoding
+                    /*.best_of   =*/ 1,
                };
            } break;
        case WHISPER_SAMPLING_BEAM_SEARCH:
            {
                result.beam_search = {
-                    /*.beam_size =*/ 2, // TODO: increase to 5 when we speed-up batch decoding
+                    /*.beam_size =*/ 5,

                    /*.patience  =*/ -1.0f,
                };
@ -3907,26 +3861,13 @@ int whisper_full_with_state(
        prompt_past.clear();
    }

-    // prepare prompt
-    {
-        std::vector<whisper_token> prompt_tokens;
-
-        // initial prompt
-        if (!params.prompt_tokens && params.initial_prompt) {
-            prompt_tokens.resize(1024);
-            prompt_tokens.resize(whisper_tokenize(ctx, params.initial_prompt, prompt_tokens.data(), prompt_tokens.size()));
-            params.prompt_tokens   = prompt_tokens.data();
-            params.prompt_n_tokens = prompt_tokens.size();
-        }
-
-        // prepend the prompt tokens to the prompt_past
-        if (params.prompt_tokens && params.prompt_n_tokens > 0) {
-            // parse tokens from the pointer
-            for (int i = 0; i < params.prompt_n_tokens; i++) {
-                prompt_past.push_back(params.prompt_tokens[i]);
-            }
-            std::rotate(prompt_past.begin(), prompt_past.end() - params.prompt_n_tokens, prompt_past.end());
+    // prepend the prompt tokens to the prompt_past
+    if (params.prompt_tokens && params.prompt_n_tokens > 0) {
+        // parse tokens from the pointer
+        for (int i = 0; i < params.prompt_n_tokens; i++) {
+            prompt_past.push_back(params.prompt_tokens[i]);
        }
+        std::rotate(prompt_past.begin(), prompt_past.end() - params.prompt_n_tokens, prompt_past.end());
    }

    // overwrite audio_ctx, max allowed is hparams.n_audio_ctx
@ -3985,10 +3926,6 @@ int whisper_full_with_state(
                fprintf(stderr, "%s: progress = %3d%%\n", __func__, progress_prev);
            }
        }
-        if (params.progress_callback) {
-            params.progress_callback(
-                ctx, ctx->state, progress_prev, params.progress_callback_user_data);
-        }

        // of only 1 second left, then stop
        if (seek + 100 >= seek_end) {
@ -4378,11 +4315,7 @@ int whisper_full_with_state(
            }

            // was the decoding successful for the current temperature?
-            // do fallback only if:
-            // - we are not at the last temperature
-            // - we are not at the end of the audio (3 sec)
-            if (it != (int) temperatures.size() - 1 &&
-                seek_end - seek > 10*WHISPER_CHUNK_SIZE) {
+            {
                bool success = true;

                const auto & decoder = state->decoders[best_decoder_id];
@ -4581,9 +4514,6 @@ int whisper_full_parallel(
        params_cur.new_segment_callback = nullptr;
        params_cur.new_segment_callback_user_data = nullptr;

-        params_cur.progress_callback = nullptr;
-        params_cur.progress_callback_user_data = nullptr;
-
        workers[i] = std::thread(whisper_full_with_state, ctx, states[i], std::move(params_cur), samples + start_samples, n_samples_cur);
    }

@ -4844,7 +4774,6 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
            struct ggml_init_params gparams = {
                /*.mem_size   =*/ buf.size(),
                /*.mem_buffer =*/ buf.data(),
-                /*.no_alloc   =*/ false,
            };

            struct ggml_context * ctx0 = ggml_init(gparams);
--- a/whisper.h
+++ b/whisper.h
@ -226,7 +226,7 @@ extern "C" {
    // Make sure to call whisper_pcm_to_mel() or whisper_set_mel() first
    // Returns the top language id or negative on failure
    // If not null, fills the lang_probs array with the probabilities of all languages
-    // The array must be whisper_lang_max_id() + 1 in size
+    // The array must be whispe_lang_max_id() + 1 in size
    // ref: https://github.com/openai/whisper/blob/main/whisper/decoding.py#L18-L69
    WHISPER_API int whisper_lang_auto_detect(
            struct whisper_context * ctx,
@ -297,7 +297,7 @@ extern "C" {

    // Available sampling strategies
    enum whisper_sampling_strategy {
-        WHISPER_SAMPLING_GREEDY,      // similar to OpenAI's GreedyDecoder
+        WHISPER_SAMPLING_GREEDY,      // similar to OpenAI's GreefyDecoder
        WHISPER_SAMPLING_BEAM_SEARCH, // similar to OpenAI's BeamSearchDecoder
    };

@ -306,9 +306,6 @@ extern "C" {
    // Use the whisper_full_...() functions to obtain the text segments
    typedef void (*whisper_new_segment_callback)(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data);

-    // Progress callback
-    typedef void (*whisper_progress_callback)(struct whisper_context * ctx, struct whisper_state * state, int progress, void * user_data);
-
    // Encoder begin callback
    // If not NULL, called before the encoder starts
    // If it returns false, the computation is aborted
@ -359,7 +356,6 @@ extern "C" {

        // tokens to provide to the whisper decoder as initial prompt
        // these are prepended to any existing text context from a previous call
-        const char * initial_prompt;
        const whisper_token * prompt_tokens;
        int prompt_n_tokens;

@ -395,10 +391,6 @@ extern "C" {
        whisper_new_segment_callback new_segment_callback;
        void * new_segment_callback_user_data;

-        // called on each progress update
-        whisper_progress_callback progress_callback;
-        void * progress_callback_user_data;
-
        // called each time before the encoder starts
        whisper_encoder_begin_callback encoder_begin_callback;
        void * encoder_begin_callback_user_data;
Author	SHA1	Message	Date
Georgi Gerganov	0244810697	rebase on master after whisper_state changes	2023-03-26 16:09:06 +03:00
Georgi Gerganov	6efb04fc72	coreml : simlpify whisper_encode + log messages	2023-03-26 15:48:45 +03:00
Georgi Gerganov	ee0d6ff473	coreml : use Core ML encoder inference	2023-03-26 15:48:41 +03:00