mirror of
https://github.com/ggerganov/whisper.cpp.git
synced 2024-12-25 15:11:05 +00:00
af5833e298
* whisper : fix cast warning * whisper : remove phase_vocoder functions, ref #2195 * whisper : remove speed_up from whisper_full_params, closes #2195
250 lines
10 KiB
C++
250 lines
10 KiB
C++
// Command line voice assisted chess
|
|
//
|
|
// Speak chess move commands to the microphone.
|
|
// The moves will translated to chessboard positions.
|
|
//
|
|
//
|
|
|
|
#include "WChess.h"
|
|
#include "common-sdl.h"
|
|
#include <iostream>
|
|
|
|
#include <memory>
|
|
#include <thread>
|
|
|
|
// command-line parameters
|
|
struct whisper_params {
|
|
int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
|
|
int32_t prompt_ms = 5000;
|
|
int32_t command_ms = 8000;
|
|
int32_t capture_id = -1;
|
|
int32_t max_tokens = 32;
|
|
int32_t audio_ctx = 0;
|
|
|
|
float vad_thold = 0.6f;
|
|
float freq_thold = 100.0f;
|
|
|
|
float grammar_penalty = 100.0f;
|
|
|
|
bool translate = false;
|
|
bool print_special = false;
|
|
bool print_energy = false;
|
|
bool no_timestamps = true;
|
|
bool use_gpu = true;
|
|
bool flash_attn = false;
|
|
|
|
std::string language = "en";
|
|
std::string model = "models/ggml-base.en.bin";
|
|
std::string fname_out;
|
|
std::string commands;
|
|
std::string prompt;
|
|
std::string context;
|
|
std::string grammar;
|
|
};
|
|
|
|
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, " -pms N, --prompt-ms N [%-7d] prompt duration in milliseconds\n", params.prompt_ms);
|
|
fprintf(stderr, " -cms N, --command-ms N [%-7d] command duration in milliseconds\n", params.command_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, " -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, " -ng, --no-gpu [%-7s] disable GPU\n", params.use_gpu ? "false" : "true");
|
|
fprintf(stderr, " -fa, --flash-attn [%-7s] flash attention during decoding\n", params.flash_attn ? "true" : "false");
|
|
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, " -cmd FNAME, --commands FNAME [%-7s] text file with allowed commands\n", params.commands.c_str());
|
|
fprintf(stderr, " -p, --prompt [%-7s] the required activation prompt\n", params.prompt.c_str());
|
|
fprintf(stderr, " -ctx, --context [%-7s] sample text to help the transcription\n", params.context.c_str());
|
|
fprintf(stderr, " --grammar-penalty N [%-7.1f] scales down logits of nongrammar tokens\n", params.grammar_penalty);
|
|
fprintf(stderr, "\n");
|
|
}
|
|
|
|
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 == "-pms" || arg == "--prompt-ms") { params.prompt_ms = std::stoi(argv[++i]); }
|
|
else if (arg == "-cms" || arg == "--command-ms") { params.command_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 == "-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 == "-ng" || arg == "--no-gpu") { params.use_gpu = false; }
|
|
else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
|
|
else if (arg == "-l" || arg == "--language") { params.language = argv[++i]; }
|
|
else if (arg == "-m" || arg == "--model") { params.model = argv[++i]; }
|
|
else if (arg == "-f" || arg == "--file") { params.fname_out = argv[++i]; }
|
|
else if (arg == "-cmd" || arg == "--commands") { params.commands = argv[++i]; }
|
|
else if (arg == "-p" || arg == "--prompt") { params.prompt = argv[++i]; }
|
|
else if (arg == "-ctx" || arg == "--context") { params.context = argv[++i]; }
|
|
else if ( arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
|
|
else {
|
|
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
|
|
whisper_print_usage(argc, argv, params);
|
|
exit(0);
|
|
}
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
std::unique_ptr<WChess> g_wchess;
|
|
int g_moveCount = 0;
|
|
void set_move(const std::string & move, float) {
|
|
if (!move.empty()) {
|
|
g_moveCount++;
|
|
fprintf(stdout, "Move: %s\n\n", move.c_str());
|
|
}
|
|
else fprintf(stdout, "Move rejected\n\n");
|
|
fprintf(stdout, "%s\n", g_wchess->stringify_board().c_str());
|
|
fprintf(stdout, "%s\n", g_moveCount ? "White's turn" : "Black's turn");
|
|
}
|
|
|
|
audio_async g_audio(30*1000);
|
|
bool g_listening = false;
|
|
std::vector<float> g_pcmf32;
|
|
|
|
bool read_input() {
|
|
std::string input;
|
|
while (true) {
|
|
fprintf(stdout, "[(l)isten/(p)ause/(q)uit]: ");
|
|
std::cin >> input;
|
|
fprintf(stdout, "\n");
|
|
if (input[0] == 'q') {
|
|
fprintf(stdout, "Quitting\n");
|
|
return false;
|
|
}
|
|
if (input[0] == 'l') {
|
|
if (!g_listening) {
|
|
fprintf(stdout, "Listening\n");
|
|
g_listening = true;
|
|
g_pcmf32.clear();
|
|
g_audio.resume();
|
|
g_audio.clear();
|
|
}
|
|
else fprintf(stdout, "Still listening\n");
|
|
return true;
|
|
}
|
|
else {
|
|
if (g_listening) {
|
|
g_listening = false;
|
|
g_audio.get(0, g_pcmf32);
|
|
g_audio.pause();
|
|
fprintf(stdout, "Processing\n");
|
|
}
|
|
else fprintf(stdout, "Not listening\n");
|
|
return true;
|
|
}
|
|
}
|
|
return true;
|
|
}
|
|
|
|
bool get_audio(std::vector<float> & pcmf32_cur) {
|
|
if (!read_input()) return false;
|
|
if (!g_pcmf32.empty()) pcmf32_cur = std::move(g_pcmf32);
|
|
else pcmf32_cur.clear();
|
|
return true;
|
|
}
|
|
|
|
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_params cparams = whisper_context_default_params();
|
|
|
|
cparams.use_gpu = params.use_gpu;
|
|
cparams.flash_attn = params.flash_attn;
|
|
|
|
struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
|
|
if (!ctx) {
|
|
fprintf(stderr, "%s: whisper_init_from_file_with_params() failed!\n", __func__);
|
|
return 1;
|
|
}
|
|
|
|
// init audio
|
|
|
|
if (!g_audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
|
|
fprintf(stderr, "%s: audio.init() failed!\n", __func__);
|
|
return 1;
|
|
}
|
|
|
|
struct whisper_full_params wparams = whisper_full_default_params(whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY);
|
|
wparams.offset_ms = 0;
|
|
wparams.translate = false;
|
|
wparams.no_context = true;
|
|
wparams.single_segment = true;
|
|
wparams.print_realtime = false;
|
|
wparams.print_progress = false;
|
|
wparams.print_timestamps = true;
|
|
wparams.print_special = false;
|
|
wparams.no_timestamps = true;
|
|
|
|
wparams.max_tokens = 32;
|
|
wparams.audio_ctx = 768; // partial encoder context for better performance
|
|
|
|
wparams.temperature = 0.0f;
|
|
wparams.temperature_inc = 2.0f;
|
|
wparams.greedy.best_of = 1;
|
|
|
|
wparams.beam_search.beam_size = 1;
|
|
|
|
wparams.language = "en";
|
|
|
|
wparams.grammar_penalty = 100.0;
|
|
|
|
wparams.initial_prompt = params.context.data();
|
|
|
|
WChess::callbacks cb;
|
|
cb.get_audio = get_audio;
|
|
cb.set_move = set_move;
|
|
|
|
WChess::settings s;
|
|
s.vad_ms = 2000;
|
|
s.prompt_ms = params.prompt_ms;
|
|
s.command_ms = params.command_ms;
|
|
s.vad_thold = params.vad_thold;
|
|
s.freq_thold = params.freq_thold;
|
|
s.print_energy = params.print_energy;
|
|
|
|
g_wchess.reset(new WChess(ctx, wparams, cb, s));
|
|
set_move("start", 0);
|
|
g_wchess->run();
|
|
|
|
whisper_print_timings(ctx);
|
|
whisper_free(ctx);
|
|
|
|
return 0;
|
|
}
|