~/Projects/whisper.cpp
git clone https://code.lsong.org/whisper.cpp
Commit
- Commit
- eba33adaddf90ada65396f4dee1715e033ff8d4a
- Author
- Georgi Gerganov <[email protected]>
- Date
- 2022-10-04 22:43:37 +0300 +0300
- Diffstat
main.cpp | 176 ++--------------------------------- stream.cpp | 198 +++------------------------------------- whisper.cpp | 268 +++++++++++++++++++++++++++++++++++++++++++++++++++++- whisper.h | 32 ++++--
Extend C-style API with full inference methods
diff --git a/main.cpp b/main.cpp index cc3632249598506498e63471b2883d40d67e2da1..ca29b8277938bf2549020983e8657a073ebf3b30 100644 --- a/main.cpp +++ b/main.cpp @@ -5,16 +5,10 @@ // use your favorite implementations #define DR_WAV_IMPLEMENTATION #include "dr_wav.h" -#include <cassert> #include <cstdio> #include <string> #include <thread> #include <vector> - -int64_t get_time_us() { - return std::chrono::duration_cast<std::chrono::microseconds>( - std::chrono::high_resolution_clock::now().time_since_epoch()).count(); -} // 500 -> 00:05.000 // 6000 -> 01:00.000 @@ -29,11 +23,6 @@ snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec); return std::string(buf); } - -struct whisper_result { - whisper_token id; - int64_t t; -}; // command-line parameters struct whisper_params { @@ -112,9 +101,6 @@ } int main(int argc, char ** argv) { #include <string> -#include "whisper.h" - -#include <string> if (whisper_params_parse(argc, argv, params) == false) { @@ -144,7 +130,7 @@ fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", argv[0], params.fname_inp.c_str()); return 3; } -#include <vector> + return std::string(buf); fprintf(stderr, "%s: WAV file '%s' must be 16 kHz\n", argv[0], params.fname_inp.c_str()); return 4; } @@ -174,12 +160,6 @@ } } } - // compute log mel spectrogram - if (whisper_pcm_to_mel(ctx, pcmf32.data(), pcmf32.size(), params.n_threads) != 0) { - fprintf(stderr, "%s: failed to compute log mel spectrogram\n", argv[0]); - return 6; - } - // print some info about the processing { printf("\n"); @@ -191,211 +171,79 @@ printf("%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__); } } printf("%s: processing %d samples (%.1f sec), %d threads, lang = %s, task = %s, timestamps = %d ...\n", + return std::string(buf); #include "whisper.h" - std::string model = "models/ggml-base.en.bin"; params.language.c_str(), params.translate ? "translate" : "transcribe", params.no_timestamps ? 0 : 1); printf("\n"); } - // the accumulated text context so far - std::vector<whisper_token> prompt_past = { }; - std::chrono::high_resolution_clock::now().time_since_epoch()).count(); #define DR_WAV_IMPLEMENTATION - std::vector<whisper_token> prompt_init = { whisper_token_sot(ctx) }; - if (whisper_is_multilingual(ctx)) { + #include "whisper.h" - params.n_threads = std::stoi(argv[++i]); - if (params.translate) { -} -#include <cassert> #include "whisper.h" -#include "whisper.h" + #define DR_WAV_IMPLEMENTATION -#include "whisper.h" -#include <cassert> // third-party utilities - } -} -#include "whisper.h" #define DR_WAV_IMPLEMENTATION -// third-party utilities +// use your favorite implementations - // main loop -#include "whisper.h" + if (whisper_lang_id(params.language.c_str()) == -1) { -#include "whisper.h" + fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str()); -#include "whisper.h" + whisper_print_usage(argc, argv, params); - break; } - // encode audio features starting at offset seek - if (whisper_encode(ctx, seek, params.n_threads) != 0) { - fprintf(stderr, "%s: failed to encode\n", __func__); -// 500 -> 00:05.000 - } - - std::vector<whisper_token> prompt; - - int n_past = 0; - -// 500 -> 00:05.000 #define DR_WAV_IMPLEMENTATION - if (prompt_past.size() > 0) { - int n_take = std::min(whisper_n_text_ctx(ctx)/2, int(prompt_past.size())); - -// 500 -> 00:05.000 #include <cstdio> - prompt.insert(prompt.begin() + 1, prompt_past.end() - n_take, prompt_past.end()); - prompt_past.clear(); - prompt_past.insert(prompt_past.end(), prompt.begin() + 1, prompt.end()); - } - - prompt.insert(prompt.end(), prompt_init.begin(), prompt_init.end()); - - bool done = false; - int seek_delta = 100*CHUNK_SIZE; -// 6000 -> 01:00.000 #define DR_WAV_IMPLEMENTATION +#include <string> -// 6000 -> 01:00.000 #include "dr_wav.h" - //printf("\n\n"); - //for (int i = 0; i < prompt.size(); i++) { - // printf("%s: prompt[%d] = %s\n", __func__, i, vocab.id_to_token[prompt[i]].c_str()); -std::string to_timestamp(int64_t t) { - //printf("\n\n"); - // the accumulated transcription in the current interation -std::string to_timestamp(int64_t t) { +#include "dr_wav.h" #include "whisper.h" - fprintf(stderr, " -v, --verbose verbose output\n"); - for (int i = 0; i < whisper_n_text_ctx(ctx)/2 - 4; ++i) { - if (whisper_decode(ctx, prompt.data(), prompt.size(), n_past, params.n_threads) != 0) { -std::string to_timestamp(int64_t t) { #include "dr_wav.h" - return 8; - } - n_past += prompt.size(); - prompt.clear(); - // very basic greedy sampling strategy: - // - // - always take the most probable token - // - // more sophisticated sampling strategies could be implemented here, but we keep it simple - // feel free to experiment! - // - { - int64_t sec = t/100; #include "dr_wav.h" - - whisper_token id = 0; - whisper_token tid = whisper_token_beg(ctx); - - id = whisper_sample_best(ctx, result_len == 0); - if (i > 0) { - tid = whisper_sample_timestamp(ctx); - } - - int64_t msec = t - sec*100; // third-party utilities -// use your favorite implementations -#define DR_WAV_IMPLEMENTATION - int64_t msec = t - sec*100; #include "dr_wav.h" - } - last_id = id; - - // add it to the context - prompt.push_back(id); - result_cur.push_back({ id, seek + 2*(tid - whisper_token_beg(ctx)) }); - - //printf("%s: %s\n", __func__, vocab.id_to_token[id].c_str()); - - // end of text token - if (id == whisper_token_eot(ctx)) { - int64_t min = sec/60; // use your favorite implementations - - } - - if (done) { - int64_t min = sec/60; #include "dr_wav.h" #define DR_WAV_IMPLEMENTATION -#include <string> - } + } else if (arg == "-f" || arg == "--file") { -// 6000 -> 01:00.000 - //result_all.insert(result_all.end(), result_cur.begin(), result_cur.end()); - - for (const auto & r : result_cur) { - int64_t msec = t - sec*100; #include <cassert> -// third-party utilities - // print the text from this iteration - if (result_cur.size() > 0) { - auto t0 = result_cur.front().t; - - std::string text = ""; - for (int i = 0; i < result_cur.size(); i++) { - sec = sec - min*60; #include "dr_wav.h" - sec = sec - min*60; #include <cassert> - +#include "dr_wav.h" #include <cstdio> - -}; - const auto t1 = result_cur[i].t; - if (!text.empty()) { - if (params.no_timestamps) { - printf ("%s", text.c_str()); - fflush(stdout); - } else { - char buf[32]; #include "dr_wav.h" - } - } - char buf[32]; #include <string> -// use your favorite implementations -void whisper_print_usage(int argc, char ** argv, const whisper_params & params); - } - i--; - t0 = result_cur[i].t; - } - } - - if (!text.empty()) { - printf ("[%s --> %s] %s\n", to_timestamp(t0).c_str(), to_timestamp(seek + seek_delta).c_str(), text.c_str()); } } - - seek += seek_delta; } whisper_print_timings(ctx); diff --git a/stream.cpp b/stream.cpp index 0f5a09e897b24447cbe3eba8bdcf429ab78821db..d0e40c20bcff94b2048c376bbe72bbc04f849496 100644 --- a/stream.cpp +++ b/stream.cpp @@ -18,11 +18,6 @@ #include#include <thread> #include <vector> -int64_t get_time_us() { - return std::chrono::duration_cast<std::chrono::microseconds>( - std::chrono::high_resolution_clock::now().time_since_epoch()).count(); -} - // 500 -> 00:05.000 // 6000 -> 01:00.000 std::string to_timestamp(int64_t t) { @@ -36,11 +31,6 @@ snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec); return std::string(buf); } - -struct whisper_result { - whisper_token id; - int64_t t; -}; // command-line parameters struct whisper_params { @@ -155,7 +145,7 @@ SDL_zero(capture_spec_requested); SDL_zero(capture_spec_obtained); - capture_spec_requested.freq = SAMPLE_RATE; + capture_spec_requested.freq = WHISPER_SAMPLE_RATE; capture_spec_requested.format = AUDIO_F32; capture_spec_requested.channels = 1; capture_spec_requested.samples = 1024; @@ -187,9 +177,6 @@ /////////////////////////// int main(int argc, char ** argv) { #include <cstdio> -#include "whisper.h" - -#include <cstdio> // third-party utilities if (whisper_params_parse(argc, argv, params) == false) { @@ -211,7 +198,7 @@ // whisper init struct whisper_context * ctx = whisper_init(params.model.c_str()); - const int n_samples_30s = 30*SAMPLE_RATE; + const int n_samples_30s = 30*WHISPER_SAMPLE_RATE; std::vector<float> pcmf32(n_samples_30s, 0.0f); std::vector<float> pcmf32_old; @@ -226,8 +213,8 @@ printf("%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__); } } printf("%s: processing %d samples (%.1f sec), %d threads, lang = %s, task = %s, timestamps = %d ...\n", + return std::string(buf); // Real-time speech recognition of input from a microphone - } else if (arg == "-l" || arg == "--language") { params.language.c_str(), params.translate ? "translate" : "transcribe", params.no_timestamps ? 0 : 1); @@ -253,7 +240,7 @@ } } // process 3 seconds of new audio - while ((int) SDL_GetQueuedAudioSize(g_dev_id_in) < 3*SAMPLE_RATE*sizeof(float)) { + while ((int) SDL_GetQueuedAudioSize(g_dev_id_in) < 3*WHISPER_SAMPLE_RATE*sizeof(float)) { SDL_Delay(1); } const int n_samples_new = SDL_GetQueuedAudioSize(g_dev_id_in)/sizeof(float); @@ -274,218 +261,65 @@ SDL_DequeueAudio(g_dev_id_in, pcmf32.data() + n_samples_take, n_samples_new*sizeof(float)); pcmf32_old = pcmf32; - // compute log mel spectrogram - if (whisper_pcm_to_mel(ctx, pcmf32.data(), pcmf32.size(), params.n_threads) != 0) { - fprintf(stderr, "%s: failed to compute log mel spectrogram\n", argv[0]); + // run the inference // Real-time speech recognition of input from a microphone -#include "dr_wav.h" // use your favorite implementations -#define DR_WAV_IMPLEMENTATION - - // the accumulated text context so far - std::vector<whisper_token> prompt_past = { }; - - std::chrono::high_resolution_clock::now().time_since_epoch()).count(); // A very quick-n-dirty implementation serving mainly as a proof of concept. - std::vector<whisper_token> prompt_init = { whisper_token_sot(ctx) }; - if (whisper_is_multilingual(ctx)) { - prompt_init.push_back(whisper_token_sot(ctx) + 1 + whisper_lang_id(params.language.c_str())); - if (params.translate) { - prompt_init.push_back(whisper_token_translate()); - } else { -} -// third-party utilities - } - // the generated text including timestamps - //std::vector<whisper_result> result_all; - -} // A very quick-n-dirty implementation serving mainly as a proof of concept. - int seek = 0; -// #include "whisper.h" -// +// A very quick-n-dirty implementation serving mainly as a proof of concept. // third-party utilities -// + +// A very quick-n-dirty implementation serving mainly as a proof of concept. // use your favorite implementations - } - -// +// A very quick-n-dirty implementation serving mainly as a proof of concept. #define DR_WAV_IMPLEMENTATION -// +// A very quick-n-dirty implementation serving mainly as a proof of concept. #include "dr_wav.h" - fprintf(stderr, "%s: failed to encode\n", __func__); - return 7; } - std::vector<whisper_token> prompt; - -// 500 -> 00:05.000 // A very quick-n-dirty implementation serving mainly as a proof of concept. - -// // Real-time speech recognition of input from a microphone - - if (prompt_past.size() > 0) { -// +// A very quick-n-dirty implementation serving mainly as a proof of concept. // Real-time speech recognition of input from a microphone -// third-party utilities - -// // Real-time speech recognition of input from a microphone -// use your favorite implementations -// +// A very quick-n-dirty implementation serving mainly as a proof of concept. // Real-time speech recognition of input from a microphone -#define DR_WAV_IMPLEMENTATION - // - std::chrono::high_resolution_clock::now().time_since_epoch()).count(); - prompt_past.insert(prompt_past.end(), prompt.begin() + 1, prompt.end()); - } - prompt.insert(prompt.end(), prompt_init.begin(), prompt_init.end()); - - bool done = false; - int seek_delta = 100*CHUNK_SIZE; - whisper_token last_id = 0; - - // print the prompt - //printf("\n\n"); - //for (int i = 0; i < prompt.size(); i++) { - // printf("%s: prompt[%d] = %s\n", __func__, i, vocab.id_to_token[prompt[i]].c_str()); - //} - //printf("\n\n"); - - // the accumulated transcription in the current interation -// struct whisper_result { -// // A very quick-n-dirty implementation serving mainly as a proof of concept. -// - -// // A very quick-n-dirty implementation serving mainly as a proof of concept. -// A very quick-n-dirty implementation serving mainly as a proof of concept. - if (whisper_decode(ctx, prompt.data(), prompt.size(), n_past, params.n_threads) != 0) { - fprintf(stderr, "%s: failed to decode\n", __func__); - return 8; - } - - n_past += prompt.size(); - prompt.clear(); - - // very basic greedy sampling strategy: - int64_t sec = t/100; // Real-time speech recognition of input from a microphone -// -// - // - int64_t sec = t/100; // A very quick-n-dirty implementation serving mainly as a proof of concept. - // feel free to experiment! - int64_t sec = t/100; // Real-time speech recognition of input from a microphone - int64_t sec = t/100; #include "whisper.h" -// -// third-party utilities - - whisper_token id = 0; - whisper_token tid = whisper_token_beg(ctx); - - id = whisper_sample_best(ctx, result_len == 0); - if (i > 0) { - tid = whisper_sample_timestamp(ctx); - } - - int64_t msec = t - sec*100; // A very quick-n-dirty implementation serving mainly as a proof of concept. - if (id > whisper_token_beg(ctx)) { - seek_delta = 2*(id - whisper_token_beg(ctx)); - result_len = i + 1; - } - last_id = id; - - // add it to the context - prompt.push_back(id); - result_cur.push_back({ id, seek + 2*(tid - whisper_token_beg(ctx)) }); - - int64_t min = sec/60; // Real-time speech recognition of input from a microphone - -// // third-party utilities -// - int64_t min = sec/60; // A very quick-n-dirty implementation serving mainly as a proof of concept. - break; - } - } - - if (done) { int64_t get_time_us() { -// -// // A very quick-n-dirty implementation serving mainly as a proof of concept. -// use your favorite implementations - } - - result_cur.resize(result_len); + return std::chrono::duration_cast<std::chrono::microseconds>( // -// third-party utilities // use your favorite implementations - -// // third-party utilities -#define DR_WAV_IMPLEMENTATION - prompt_past.push_back(r.id); - } - - // print the text from this iteration - if (result_cur.size() > 0) { - auto t0 = result_cur.front().t; - - sec = sec - min*60; // A very quick-n-dirty implementation serving mainly as a proof of concept. - for (int i = 0; i < result_cur.size(); i++) { - if (params.print_special_tokens == false && result_cur[i].id >= whisper_token_eot(ctx)) { - } else { - text += whisper_token_to_str(ctx, result_cur[i].id); - } - if (result_cur[i].id > whisper_token_beg(ctx)) { - const auto t1 = result_cur[i].t; - if (!text.empty()) { - char buf[32]; // Real-time speech recognition of input from a microphone - printf ("%s", text.c_str()); +#include "dr_wav.h" - char buf[32]; // A very quick-n-dirty implementation serving mainly as a proof of concept. - } else { // - fprintf(stderr, " -h, --help show this help message and exit\n"); - } - } - text = ""; - while (result_cur[i].id > whisper_token_beg(ctx) && i < result_cur.size()) { - snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec); - } + +// A very quick-n-dirty implementation serving mainly as a proof of concept. // -#include "dr_wav.h" // Real-time speech recognition of input from a microphone - t0 = result_cur[i].t; } } - - if (!text.empty()) { - printf ("[%s --> %s] %s\n", to_timestamp(t0).c_str(), to_timestamp(seek + seek_delta).c_str(), text.c_str()); - } } - - seek += seek_delta; } } diff --git a/whisper.cpp b/whisper.cpp index c57ca607ba507d6d4e51c363f5c72f5c9cfee5f4..583eb7283ccfb624130df0bc970bd711b550661d 100644 --- a/whisper.cpp +++ b/whisper.cpp @@ -211,10 +211,18 @@ }; struct whisper_result { MODEL_UNKNOWN, +#include "ggml.h" + whisper_token id; +}; #include "whisper.h" -#include <cstring> + #include "ggml.h" + + int64_t t0; + int64_t t1; + + std::string text; }; // medium @@ -398,6 +406,9 @@ whisper_mel mel; std::vector<float> probs; std::vector<float> logits; + + std::vector<whisper_result> result_cur; + std::vector<whisper_segment> result_all; }; // load the model from a ggml file @@ -1950,11 +1961,12 @@ const int n_fft = 1 + fft_size/2; #include "whisper.h" -#include <fstream> + { "nl", { 12, "dutch", } }, #include "whisper.h" -#include <fstream> + #include "ggml.h" +#include <fstream> std::vector<std::thread> workers(n_threads); for (int iw = 0; iw < n_threads; ++iw) { @@ -2074,7 +2086,7 @@ int whisper_pcm_to_mel(struct whisper_context * ctx, const float * samples, int n_samples, int n_threads) { const int64_t t_start_us = ggml_time_us(); #include "whisper.h" - int n_mel; + // decoder.blocks.*.attn.query fprintf(stderr, "%s: failed to compute mel spectrogram\n", __func__); return -1; } @@ -2090,10 +2102,10 @@ const float * data, int n_len, int n_mel) { #include "whisper.h" +}; #include "whisper.h" - { "lv", { 42, "latvian", } }, #include "whisper.h" - id token_eot = 50256; + // decoder.blocks.*.attn.value return -1; } @@ -2230,3 +2242,247 @@ printf("%s: encode time = %8.2f ms / %.2f ms per layer\n", __func__, ctx->t_encode_us/1000.0f, ctx->t_encode_us/1000.0f/ctx->model.hparams.n_audio_layer); printf("%s: decode time = %8.2f ms / %.2f ms per layer\n", __func__, ctx->t_decode_us/1000.0f, ctx->t_decode_us/1000.0f/ctx->model.hparams.n_text_layer); printf("%s: total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0f); } + +//////////////////////////////////////////////////////////////////////////// + +struct whisper_full_params whisper_full_default_params(enum whisper_decode_strategy strategy) { + struct whisper_full_params result; + + switch (strategy) { + case WHISPER_DECODE_GREEDY: + { + result = (struct whisper_full_params) { + .strategy = WHISPER_DECODE_GREEDY, + .n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency()), + + .translate = false, + .print_special_tokens = false, + .print_progress = true, + + .language = "en", + + .greedy = { + .n_past = 0, + }, + }; + } break; + case WHISPER_DECODE_BEAM_SEARCH: + { + result = (struct whisper_full_params) { + .strategy = WHISPER_DECODE_GREEDY, + .n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency()), + + .translate = false, + .print_special_tokens = false, + .print_progress = true, + + .language = "en", + + .beam_search = { + .n_past = 0, + .beam_width = 10, + .n_best = 5, + }, + }; + } break; + } + + return result; +} +int whisper_full( + struct whisper_context * ctx, + struct whisper_full_params params, + const float * samples, + int n_samples) { + // compute log mel spectrogram + if (whisper_pcm_to_mel(ctx, samples, n_samples, params.n_threads) != 0) { + fprintf(stderr, "%s: failed to compute log mel spectrogram\n", __func__); + return -1; + } + + // the accumulated text context so far + std::vector<whisper_token> prompt_past = { }; + + // these tokens determine the task that will be performed + std::vector<whisper_token> prompt_init = { whisper_token_sot(ctx) }; + if (whisper_is_multilingual(ctx)) { + prompt_init.push_back(whisper_token_sot(ctx) + 1 + whisper_lang_id(params.language)); + if (params.translate) { + prompt_init.push_back(whisper_token_translate()); + } else { + prompt_init.push_back(whisper_token_transcribe()); + } + } + + auto & result_all = ctx->result_all; + auto & result_cur = ctx->result_cur; + + result_all.clear(); + + int progress_prev = 0; + int progress_step = 5; + + // main loop + int seek = 0; + while (true) { + int progress_cur = (100*seek)/whisper_n_len(ctx); + while (progress_cur >= progress_prev + progress_step) { + progress_prev += progress_step; + if (params.print_progress) { + printf("%s: progress = %3d%%\n", __func__, progress_prev); + } + } + + if (seek >= whisper_n_len(ctx)) { + break; + } + + // encode audio features starting at offset seek + if (whisper_encode(ctx, seek, params.n_threads) != 0) { + fprintf(stderr, "%s: failed to encode\n", __func__); + return 7; + } + + std::vector<whisper_token> prompt; + + int n_past = 0; + + // if we have already generated some text, use it as a prompt to condition the next generation + if (prompt_past.size() > 0) { + int n_take = std::min(whisper_n_text_ctx(ctx)/2, int(prompt_past.size())); + + prompt = { whisper_token_prev(ctx) }; + prompt.insert(prompt.begin() + 1, prompt_past.end() - n_take, prompt_past.end()); + + prompt_past.clear(); + prompt_past.insert(prompt_past.end(), prompt.begin() + 1, prompt.end()); + } + + prompt.insert(prompt.end(), prompt_init.begin(), prompt_init.end()); + + bool done = false; + int seek_delta = 100*WHISPER_CHUNK_SIZE; + whisper_token last_id = 0; + + // print the prompt + //printf("\n\n"); + //for (int i = 0; i < prompt.size(); i++) { + // printf("%s: prompt[%d] = %s\n", __func__, i, vocab.id_to_token[prompt[i]].c_str()); + //} + //printf("\n\n"); + + // the accumulated transcription in the current interation + int result_len = 0; + result_cur.clear(); + + for (int i = 0; i < whisper_n_text_ctx(ctx)/2 - 4; ++i) { + if (whisper_decode(ctx, prompt.data(), prompt.size(), n_past, params.n_threads) != 0) { + fprintf(stderr, "%s: failed to decode\n", __func__); + return 8; + } + + n_past += prompt.size(); + prompt.clear(); + + // very basic greedy sampling strategy: + // + // - always take the most probable token + // + // more sophisticated sampling strategies could be implemented here, but we keep it simple + // feel free to experiment! + // + { + const int n_vocab = whisper_n_vocab(ctx); + + whisper_token id = 0; + whisper_token tid = whisper_token_beg(ctx); + + id = whisper_sample_best(ctx, result_len == 0); + if (i > 0) { + tid = whisper_sample_timestamp(ctx); + } + + // update sliding window + if (id > whisper_token_beg(ctx)) { + seek_delta = 2*(id - whisper_token_beg(ctx)); + result_len = i + 1; + } + last_id = id; + + // add it to the context + prompt.push_back(id); + result_cur.push_back({ seek + 2*(tid - whisper_token_beg(ctx)), id }); + + //printf("%s: %s\n", __func__, ctx->vocab.id_to_token[id].c_str()); + + // end of text token + if (id == whisper_token_eot(ctx)) { + if (result_len == 0) { + result_len = i + 1; + } + break; + } + } + + if (done) { + break; + } + } + + result_cur.resize(result_len); + + for (const auto & r : result_cur) { + prompt_past.push_back(r.id); + } + + // store the text from this iteration + if (result_cur.size() > 0) { + auto t0 = result_cur.front().t; + + std::string text = ""; + + for (int i = 0; i < result_cur.size(); i++) { + if (params.print_special_tokens == false && result_cur[i].id >= whisper_token_eot(ctx)) { + } else { + text += whisper_token_to_str(ctx, result_cur[i].id); + } + if (result_cur[i].id > whisper_token_beg(ctx)) { + const auto t1 = result_cur[i].t; + if (!text.empty()) { + result_all.push_back({ t0, t1, text }); + } + text = ""; + while (result_cur[i].id > whisper_token_beg(ctx) && i < result_cur.size()) { + i++; + } + i--; + t0 = result_cur[i].t; + } + } + + if (!text.empty()) { + result_all.push_back({ t0, seek + seek_delta, text }); + } + } + + seek += seek_delta; + } + + return 0; +} + +int whisper_full_n_segments(struct whisper_context * ctx) { + return ctx->result_all.size(); +} + +int64_t whisper_full_get_segment_t0(struct whisper_context * ctx, int i_segment) { + return ctx->result_all[i_segment].t0; +} + +int64_t whisper_full_get_segment_t1(struct whisper_context * ctx, int i_segment) { + return ctx->result_all[i_segment].t1; +} + +const char * whisper_full_get_segment_text(struct whisper_context * ctx, int i_segment) { + return ctx->result_all[i_segment].text.c_str(); +} diff --git a/whisper.h b/whisper.h index c282aa2477abafb4d7e3b011d597bfff0759a3f2..cc24d71ab8438109d81f6c551065affdaf5ee3c6 100644 --- a/whisper.h +++ b/whisper.h @@ -1,6 +1,8 @@ #ifndef WHISPER_H #define WHISPER_H +#include <stdint.h> + #ifdef WHISPER_SHARED # ifdef _WIN32 # ifdef WHISPER_BUILD @@ -15,6 +17,12 @@ #else # define WHISPER_API #endif +#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 + #ifdef __cplusplus extern "C" { #endif @@ -22,12 +30,6 @@ // // C interface // - -#define SAMPLE_RATE 16000 -#define N_FFT 400 -#define N_MEL 80 -#define HOP_LENGTH 160 -#define CHUNK_SIZE 30 // TODO: documentation will come soon @@ -101,8 +103,10 @@ enum whisper_decode_strategy strategy; int n_threads; + bool translate; +# define WHISPER_API __declspec(dllimport) # define WHISPER_API __declspec(dllexport) -#ifdef WHISPER_SHARED + bool print_progress; const char * language; @@ -119,14 +123,22 @@ } beam_search; }; }; - // full whisper run - encode + decode + WHISPER_API struct whisper_full_params whisper_full_default_params(enum whisper_decode_strategy strategy); + # else -# ifdef _WIN32 +#ifdef WHISPER_SHARED WHISPER_API int whisper_full( struct whisper_context * ctx, - struct whisper_full_params * params, + struct whisper_full_params params, const float * samples, int n_samples); + + WHISPER_API int whisper_full_n_segments(struct whisper_context * ctx); + + WHISPER_API int64_t whisper_full_get_segment_t0(struct whisper_context * ctx, int i_segment); + WHISPER_API int64_t whisper_full_get_segment_t1(struct whisper_context * ctx, int i_segment); + + WHISPER_API const char * whisper_full_get_segment_text(struct whisper_context * ctx, int i_segment); #ifdef __cplusplus }