whisper : add support for large v3 (#1444)

* whisper : add support for large v3

* bench : fix build + fix go bindings

* bench : fix n_mels

* models : update readme

Makefile

@@ -417,9 +417,10 @@ samples:
 .PHONY: medium.en
 .PHONY: medium
 .PHONY: large-v1
+.PHONY: large-v2
 .PHONY: large
 
-tiny.en tiny base.en base small.en small medium.en medium large-v1 large: main
+tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large: main
 	bash ./models/download-ggml-model.sh $@
 	@echo ""
 	@echo "==============================================="

README.md

@@ -234,6 +234,7 @@ make small
 make medium.en
 make medium
 make large-v1
+make large-v2
 make large
 ```
 
@@ -245,7 +246,7 @@ make large
 | base   | 142 MB | ~210 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
 | small  | 466 MB | ~600 MB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
 | medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
-| large  | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
+| large  | 2.9 GB | ~3.3 GB | `ad82bf6a9043ceed055076d0fd39f5f186ff8062` |
 
 ## Quantization
 

bindings/go/examples/go-model-download/main.go

@@ -24,7 +24,7 @@ const (
 
 var (
 	// The models which will be downloaded, if no model is specified as an argument
-	modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large"}
+	modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large-v2", "ggml-large"}
 )
 
 var (

bindings/go/whisper.go

@@ -83,7 +83,6 @@ const (
 	SampleRate = C.WHISPER_SAMPLE_RATE                 // Expected sample rate, samples per second
 	SampleBits = uint16(unsafe.Sizeof(C.float(0))) * 8 // Sample size in bits
 	NumFFT     = C.WHISPER_N_FFT
-	NumMEL     = C.WHISPER_N_MEL
 	HopLength  = C.WHISPER_HOP_LENGTH
 	ChunkSize  = C.WHISPER_CHUNK_SIZE
 )

examples/bench.wasm/emscripten.cpp

@@ -23,7 +23,9 @@ void bench_main(size_t index) {
 
     fprintf(stderr, "%s: running benchmark with %d threads - please wait...\n", __func__, n_threads);
 
-    if (int ret = whisper_set_mel(ctx, nullptr, 0, WHISPER_N_MEL)) {
+    const int n_mels = whisper_model_n_mels(ctx);
+
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
         fprintf(stderr, "error: failed to set mel: %d\n", ret);
         return;
     }

examples/bench/bench.cpp

@@ -73,7 +73,9 @@ int whisper_bench_full(const whisper_params & params) {
         return 2;
     }
 
-    if (int ret = whisper_set_mel(ctx, nullptr, 0, WHISPER_N_MEL)) {
+    const int n_mels = whisper_model_n_mels(ctx);
+
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
         fprintf(stderr, "error: failed to set mel: %d\n", ret);
         return 3;
     }

examples/livestream.sh

@@ -48,7 +48,7 @@ if [ -n "$3" ]; then
 fi
 
 # Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large" )
 
 # list available models
 function list_models {

examples/twitch.sh

@@ -21,7 +21,7 @@ help()
     echo "Usage: ./twitch.sh -s [step] -m [model] -t [threads] [url]"
     echo "options:"
     echo "-s       Step in seconds (default is $step)."
-    echo "-m       Choose model, options are: 'tiny.en' 'tiny' 'base.en' 'base' 'small.en' 'small' 'medium.en' 'medium' 'large-v1' 'large' (default is '$model')."
+    echo "-m       Choose model, options are: 'tiny.en' 'tiny' 'base.en' 'base' 'small.en' 'small' 'medium.en' 'medium' 'large-v1' 'large-v2' 'large' (default is '$model')."
     echo "-t       Number of threads to use."
     echo "-h       Print this help page."
     echo

extra/convert-all.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large" )
 
 for model in "${models[@]}"; do
     python3 models/convert-pt-to-ggml.py ~/.cache/whisper/$model.pt ../whisper models/

models/README.md

@@ -50,7 +50,8 @@ https://huggingface.co/ggerganov/whisper.cpp/tree/main
 | medium    | 1.5 GB | ~2.6 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
 | medium.en | 1.5 GB | ~2.6 GB | `8c30f0e44ce9560643ebd10bbe50cd20eafd3723` |
 | large-v1  | 2.9 GB | ~4.7 GB | `b1caaf735c4cc1429223d5a74f0f4d0b9b59a299` |
-| large     | 2.9 GB | ~4.7 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
+| large-v2  | 2.9 GB | ~4.7 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
+| large     | 2.9 GB | ~4.7 GB | `ad82bf6a9043ceed055076d0fd39f5f186ff8062` |
 
 ## Model files for testing purposes
 

models/convert-h5-to-coreml.py

@@ -78,14 +78,14 @@ def convert_hf_whisper(hf_model_name_or_path: str, whisper_state_path: str):
 # Ported from models/convert-whisper-to-coreml.py
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model-name", type=str, help="name of model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large, large-v1)", required=True)
+    parser.add_argument("--model-name", type=str, help="name of model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large, large-v1, large-v2)", required=True)
     parser.add_argument("--model-path", type=str, help="path to the model (e.g. if published on HuggingFace: Oblivion208/whisper-tiny-cantonese)", 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_name not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large", "large-v1"]:
+    if args.model_name not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large", "large-v1", "large-v2"]:
         raise ValueError("Invalid model name")
 
     pt_target_path = f"models/hf-{args.model_name}.pt"

models/convert-pt-to-ggml.py

@@ -228,7 +228,7 @@ with np.load(dir_whisper / "whisper" / "assets" / "mel_filters.npz") as f:
 # for backwards compatibility, also check for older hf_transformers format tokenizer files
 # old format: dir_whisper/whisper/assets/[multilingual/gpt2]/vocab.json
 # new format: dir_whisper/whisper/assets/[multilingual/gpt2].tiktoken
-multilingual = hparams["n_vocab"] == 51865
+multilingual = hparams["n_vocab"] >= 51865
 tokenizer = dir_whisper / "whisper" / "assets" / (multilingual and "multilingual.tiktoken" or "gpt2.tiktoken")
 tokenizer_type = "tiktoken"
 if not tokenizer.is_file():

models/convert-whisper-to-coreml.py

@@ -194,7 +194,7 @@ class TextDecoderANE(TextDecoder):
         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:
+        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)
@@ -296,13 +296,13 @@ def convert_decoder(hparams, model, quantize=False):
 
 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, large-v1)", required=True)
+    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, large-v1, large-v2)", 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", "large-v1"]:
+    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large", "large-v1", "large-v2"]:
         raise ValueError("Invalid model name")
 
     whisper = load_model(args.model).cpu()

models/convert-whisper-to-openvino.py

@@ -38,10 +38,10 @@ def convert_encoder(hparams, encoder, mname):
 
 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, large-v1)", required=True)
+    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, large-v1, large-v2)", required=True)
     args = parser.parse_args()
 
-    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large", "large-v1"]:
+    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large", "large-v1", "large-v2"]:
         raise ValueError("Invalid model name")
 
     whisper = load_model(args.model).cpu()

models/download-coreml-model.sh

@@ -19,7 +19,7 @@ function get_script_path() {
 models_path="$(get_script_path)"
 
 # Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large" )
 
 # list available models
 function list_models {

models/download-ggml-model.cmd

@@ -8,7 +8,7 @@ popd
 set argc=0
 for %%x in (%*) do set /A argc+=1
 
-set models=tiny.en tiny base.en base small.en small medium.en medium large-v1 large
+set models=tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large
 
 if %argc% neq 1 (
   echo.

models/download-ggml-model.sh

@@ -41,6 +41,7 @@ models=(
     "medium-q5_0"
     "medium.en-q5_0"
     "large-v1"
+    "large-v2"
     "large"
     "large-q5_0"
 )

tests/run-tests.sh

@@ -19,7 +19,7 @@
 cd `dirname $0`
 
 # Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large" )
 
 # list available models
 function list_models {

whisper.cpp

@@ -193,6 +193,15 @@ enum e_model {
     MODEL_LARGE,
 };
 
+static const std::map<e_model, std::string> g_model_name = {
+    { MODEL_UNKNOWN,  "unknown"  },
+    { MODEL_TINY,     "tiny"     },
+    { MODEL_BASE,     "base"     },
+    { MODEL_SMALL,    "small"    },
+    { MODEL_MEDIUM,   "medium"   },
+    { MODEL_LARGE,    "large"    },
+};
+
 static const std::map<std::string, std::pair<int, std::string>> g_lang = {
     { "en",  { 0,  "english",         } },
     { "zh",  { 1,  "chinese",         } },
@@ -293,6 +302,7 @@ static const std::map<std::string, std::pair<int, std::string>> g_lang = {
     { "ba",  { 96,  "bashkir",        } },
     { "jw",  { 97,  "javanese",       } },
     { "su",  { 98,  "sundanese",      } },
+    { "yue", { 99,  "cantonese",      } },
 };
 
 static const size_t MB = 1ull*1024*1024;
@@ -402,7 +412,11 @@ struct whisper_vocab {
     id token_beg        = 50363; // begin timestamps
 
     bool is_multilingual() const {
-        return n_vocab == 51865;
+        return n_vocab >= 51865;
+    }
+
+    int num_languages() const {
+        return n_vocab - 51765 - (is_multilingual() ? 1 : 0);
     }
 };
 
@@ -922,6 +936,8 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
 
         assert(hparams.n_text_state == hparams.n_audio_state);
 
+        std::string mver = "";
+
         if (hparams.n_audio_layer == 4) {
             model.type = e_model::MODEL_TINY;
         }
@@ -940,6 +956,10 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
 
         if (hparams.n_audio_layer == 32) {
             model.type = e_model::MODEL_LARGE;
+
+            if (hparams.n_vocab == 51866) {
+                mver = " v3";
+            }
         }
 
         const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
@@ -968,7 +988,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
         log("%s: n_mels        = %d\n", __func__, hparams.n_mels);
         log("%s: ftype         = %d\n", __func__, model.hparams.ftype);
         log("%s: qntvr         = %d\n", __func__, qntvr);
-        log("%s: type          = %d\n", __func__, model.type);
+        log("%s: type          = %d (%s%s)\n", __func__, model.type, g_model_name.at(model.type).c_str(), mver.c_str());
 
         // print memory requirements
         {
@@ -1039,13 +1059,17 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
         if (vocab.is_multilingual()) {
             vocab.token_eot++;
             vocab.token_sot++;
-            vocab.token_translate++;
+
-            vocab.token_transcribe++;
+            // account for variable number of language tokens
-            vocab.token_solm++;
+            const int dt = vocab.num_languages() - 98;
-            vocab.token_prev++;
+
-            vocab.token_nosp++;
+            vocab.token_translate  += dt;
-            vocab.token_not++;
+            vocab.token_transcribe += dt;
-            vocab.token_beg++;
+            vocab.token_solm       += dt;
+            vocab.token_prev       += dt;
+            vocab.token_nosp       += dt;
+            vocab.token_not        += dt;
+            vocab.token_beg        += dt;
         }
 
         if (n_vocab < model.hparams.n_vocab) {
@@ -1074,6 +1098,8 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
                 vocab.id_to_token[i] = word;
             }
         }
+
+        log("%s: n_langs       = %d\n", __func__, vocab.num_languages());
     }
 
     size_t ctx_size = 0;
@@ -3281,7 +3307,7 @@ void whisper_free_params(struct whisper_full_params * params) {
 }
 
 int whisper_pcm_to_mel_with_state(struct whisper_context * ctx, struct whisper_state * state, const float * samples, int n_samples, int n_threads) {
-    if (!log_mel_spectrogram(*state, samples, n_samples, WHISPER_SAMPLE_RATE, WHISPER_N_FFT, WHISPER_HOP_LENGTH, WHISPER_N_MEL, n_threads, ctx->model.filters, false, state->mel)) {
+    if (!log_mel_spectrogram(*state, samples, n_samples, WHISPER_SAMPLE_RATE, WHISPER_N_FFT, WHISPER_HOP_LENGTH, ctx->model.filters.n_mel, n_threads, ctx->model.filters, false, state->mel)) {
         log("%s: failed to compute mel spectrogram\n", __func__);
         return -1;
     }
@@ -3295,7 +3321,7 @@ int whisper_pcm_to_mel(struct whisper_context * ctx, const float * samples, int
 
 // same as whisper_pcm_to_mel, but applies a Phase Vocoder to speed up the audio x2 (PV without phase lock is not good)
 int whisper_pcm_to_mel_phase_vocoder_with_state(struct whisper_context * ctx, struct whisper_state * state, const float * samples, int n_samples, int n_threads) {
-    if (!log_mel_spectrogram(*state, samples, n_samples, WHISPER_SAMPLE_RATE, 2 * WHISPER_N_FFT, 2 * WHISPER_HOP_LENGTH, WHISPER_N_MEL, n_threads, ctx->model.filters, false, state->mel)) {
+    if (!log_mel_spectrogram(*state, samples, n_samples, WHISPER_SAMPLE_RATE, 2 * WHISPER_N_FFT, 2 * WHISPER_HOP_LENGTH, ctx->model.filters.n_mel, n_threads, ctx->model.filters, false, state->mel)) {
         log("%s: failed to compute mel spectrogram\n", __func__);
         return -1;
     }
@@ -3318,13 +3344,13 @@ int whisper_pcm_to_mel_phase_vocoder(struct whisper_context * ctx, const float *
 // TODO
 
 int whisper_set_mel_with_state(
-        struct whisper_context * /*ctx*/,
+        struct whisper_context * ctx,
           struct whisper_state * state,
                    const float * data,
                            int   n_len,
                            int   n_mel) {
-    if (n_mel != WHISPER_N_MEL) {
+    if (n_mel != ctx->model.filters.n_mel) {
-        log("%s: invalid number of mel bands: %d (expected %d)\n", __func__, n_mel, WHISPER_N_MEL);
+        log("%s: invalid number of mel bands: %d (expected %d)\n", __func__, n_mel, ctx->model.filters.n_mel);
         return -1;
     }
 

whisper.h

@@ -29,7 +29,6 @@
 
 #define WHISPER_SAMPLE_RATE 16000
 #define WHISPER_N_FFT       400
-#define WHISPER_N_MEL       80
 #define WHISPER_HOP_LENGTH  160
 #define WHISPER_CHUNK_SIZE  30