diff --git a/examples/stream/stream.cpp b/examples/stream/stream.cpp index 92fee4e..d281609 100644 --- a/examples/stream/stream.cpp +++ b/examples/stream/stream.cpp @@ -1,6 +1,7 @@ // Real-time speech recognition of input from a microphone // // A very quick-n-dirty implementation serving mainly as a proof of concept. +// #include "whisper.h" @@ -33,15 +34,19 @@ struct whisper_params { int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency()); int32_t step_ms = 3000; int32_t length_ms = 10000; + int32_t keep_ms = 200; 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; + bool speed_up = false; bool translate = false; - bool no_context = true; bool print_special = false; - bool no_timestamps = true; + bool no_context = true; + bool no_timestamps = false; std::string language = "en"; std::string model = "models/ggml-base.en.bin"; @@ -61,13 +66,16 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) { else if (arg == "-t" || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); } else if ( arg == "--step") { params.step_ms = std::stoi(argv[++i]); } else if ( arg == "--length") { params.length_ms = std::stoi(argv[++i]); } + else if ( arg == "--keep") { params.keep_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 == "-su" || arg == "--speed-up") { params.speed_up = true; } else if (arg == "-tr" || arg == "--translate") { params.translate = true; } - else if (arg == "-kc" || arg == "--keep-context") { params.no_context = false; } 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]; } else if (arg == "-m" || arg == "--model") { params.model = argv[++i]; } else if (arg == "-f" || arg == "--file") { params.fname_out = argv[++i]; } @@ -90,13 +98,16 @@ void whisper_print_usage(int argc, char ** argv, const whisper_params & params) 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, " -kc, --keep-context [%-7s] keep context between audio chunks\n", params.no_context ? "false" : "true"); 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()); @@ -107,19 +118,56 @@ void whisper_print_usage(int argc, char ** argv, const whisper_params & params) // SDL Audio capture // -SDL_AudioDeviceID g_dev_id_in = 0; +class audio_async { +public: + audio_async(int len_ms); + ~audio_async(); -bool audio_sdl_init(const int capture_id) { - if (g_dev_id_in) { - fprintf(stderr, "%s: already initialized\n", __func__); - return false; + bool init(int capture_id, int sample_rate); + + // start capturing audio via the provided SDL callback + // keep last len_ms seconds of audio in a circular buffer + bool resume(); + bool pause(); + bool clear(); + + // callback to be called by SDL + void callback(uint8_t * stream, int len); + + // get audio data from the circular buffer + void get(int ms, std::vector & audio); + +private: + SDL_AudioDeviceID m_dev_id_in = 0; + + int m_len_ms = 0; + int m_sample_rate = 0; + + bool m_running = false; + std::mutex m_mutex; + + std::vector m_audio; + std::vector m_audio_new; + size_t m_audio_pos = 0; + size_t m_audio_len = 0; +}; + +audio_async::audio_async(int len_ms) { + m_len_ms = len_ms; +} + +audio_async::~audio_async() { + if (m_dev_id_in) { + SDL_CloseAudioDevice(m_dev_id_in); } +} +bool audio_async::init(int capture_id, int sample_rate) { SDL_LogSetPriority(SDL_LOG_CATEGORY_APPLICATION, SDL_LOG_PRIORITY_INFO); if (SDL_Init(SDL_INIT_AUDIO) < 0) { SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "Couldn't initialize SDL: %s\n", SDL_GetError()); - return (1); + return false; } SDL_SetHintWithPriority(SDL_HINT_AUDIO_RESAMPLING_MODE, "medium", SDL_HINT_OVERRIDE); @@ -138,34 +186,232 @@ bool audio_sdl_init(const int capture_id) { SDL_zero(capture_spec_requested); SDL_zero(capture_spec_obtained); - capture_spec_requested.freq = WHISPER_SAMPLE_RATE; + capture_spec_requested.freq = sample_rate; capture_spec_requested.format = AUDIO_F32; capture_spec_requested.channels = 1; capture_spec_requested.samples = 1024; + capture_spec_requested.callback = [](void * userdata, uint8_t * stream, int len) { + audio_async * audio = (audio_async *) userdata; + audio->callback(stream, len); + }; + capture_spec_requested.userdata = this; if (capture_id >= 0) { fprintf(stderr, "%s: attempt to open capture device %d : '%s' ...\n", __func__, capture_id, SDL_GetAudioDeviceName(capture_id, SDL_TRUE)); - g_dev_id_in = SDL_OpenAudioDevice(SDL_GetAudioDeviceName(capture_id, SDL_TRUE), SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0); + m_dev_id_in = SDL_OpenAudioDevice(SDL_GetAudioDeviceName(capture_id, SDL_TRUE), SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0); } else { fprintf(stderr, "%s: attempt to open default capture device ...\n", __func__); - g_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0); + m_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0); } - if (!g_dev_id_in) { + + if (!m_dev_id_in) { fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError()); - g_dev_id_in = 0; + m_dev_id_in = 0; + + return false; } else { - fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, g_dev_id_in); - fprintf(stderr, "%s: - sample rate: %d\n", __func__, capture_spec_obtained.freq); - fprintf(stderr, "%s: - format: %d (required: %d)\n", __func__, capture_spec_obtained.format, capture_spec_requested.format); - fprintf(stderr, "%s: - channels: %d (required: %d)\n", __func__, capture_spec_obtained.channels, capture_spec_requested.channels); - fprintf(stderr, "%s: - samples per frame: %d\n", __func__, capture_spec_obtained.samples); + fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, m_dev_id_in); + fprintf(stderr, "%s: - sample rate: %d\n", __func__, capture_spec_obtained.freq); + fprintf(stderr, "%s: - format: %d (required: %d)\n", __func__, capture_spec_obtained.format, + capture_spec_requested.format); + fprintf(stderr, "%s: - channels: %d (required: %d)\n", __func__, capture_spec_obtained.channels, + capture_spec_requested.channels); + fprintf(stderr, "%s: - samples per frame: %d\n", __func__, capture_spec_obtained.samples); + } + + m_sample_rate = capture_spec_obtained.freq; + + m_audio.resize((m_sample_rate*m_len_ms)/1000); + + return true; +} + +bool audio_async::resume() { + if (!m_dev_id_in) { + fprintf(stderr, "%s: no audio device to resume!\n", __func__); + return false; + } + + if (m_running) { + fprintf(stderr, "%s: already running!\n", __func__); + return false; + } + + SDL_PauseAudioDevice(m_dev_id_in, 0); + + m_running = true; + + return true; +} + +bool audio_async::pause() { + if (!m_dev_id_in) { + fprintf(stderr, "%s: no audio device to pause!\n", __func__); + return false; + } + + if (!m_running) { + fprintf(stderr, "%s: already paused!\n", __func__); + return false; + } + + SDL_PauseAudioDevice(m_dev_id_in, 1); + + m_running = false; + + return true; +} + +bool audio_async::clear() { + if (!m_dev_id_in) { + fprintf(stderr, "%s: no audio device to clear!\n", __func__); + return false; + } + + if (!m_running) { + fprintf(stderr, "%s: not running!\n", __func__); + return false; + } + + { + std::lock_guard lock(m_mutex); + + m_audio_pos = 0; + m_audio_len = 0; } return true; } +// callback to be called by SDL +void audio_async::callback(uint8_t * stream, int len) { + if (!m_running) { + return; + } + + const size_t n_samples = len / sizeof(float); + + m_audio_new.resize(n_samples); + memcpy(m_audio_new.data(), stream, n_samples * sizeof(float)); + + //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len); + + { + std::lock_guard lock(m_mutex); + + if (m_audio_pos + n_samples > m_audio.size()) { + const size_t n0 = m_audio.size() - m_audio_pos; + + memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float)); + memcpy(&m_audio[0], &stream[n0], (n_samples - n0) * sizeof(float)); + + m_audio_pos = (m_audio_pos + n_samples) % m_audio.size(); + m_audio_len = m_audio.size(); + } else { + memcpy(&m_audio[m_audio_pos], stream, n_samples * sizeof(float)); + + m_audio_pos = (m_audio_pos + n_samples) % m_audio.size(); + m_audio_len = std::min(m_audio_len + n_samples, m_audio.size()); + } + } +} + +void audio_async::get(int ms, std::vector & result) { + if (!m_dev_id_in) { + fprintf(stderr, "%s: no audio device to get audio from!\n", __func__); + return; + } + + if (!m_running) { + fprintf(stderr, "%s: not running!\n", __func__); + return; + } + + result.clear(); + + { + std::lock_guard lock(m_mutex); + + if (ms <= 0) { + ms = m_len_ms; + } + + size_t n_samples = (m_sample_rate * ms) / 1000; + if (n_samples > m_audio_len) { + n_samples = m_audio_len; + } + + result.resize(n_samples); + + int s0 = m_audio_pos - n_samples; + if (s0 < 0) { + s0 += m_audio.size(); + } + + if (s0 + n_samples > m_audio.size()) { + const size_t n0 = m_audio.size() - s0; + + memcpy(result.data(), &m_audio[s0], n0 * sizeof(float)); + memcpy(&result[n0], &m_audio[0], (n_samples - n0) * sizeof(float)); + } else { + memcpy(result.data(), &m_audio[s0], n_samples * sizeof(float)); + } + } +} + /////////////////////////// +void high_pass_filter(std::vector & data, float cutoff, float sample_rate) { + const float rc = 1.0f / (2.0f * M_PI * cutoff); + const float dt = 1.0f / sample_rate; + const float alpha = dt / (rc + dt); + + float y = data[0]; + + for (size_t i = 1; i < data.size(); i++) { + y = alpha * (y + data[i] - data[i - 1]); + data[i] = y; + } +} + +bool vad_simple(std::vector & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) { + const int n_samples = pcmf32.size(); + const int n_samples_last = (sample_rate * last_ms) / 1000; + + if (n_samples_last >= n_samples) { + // not enough samples - assume no speech + return false; + } + + if (freq_thold > 0.0f) { + high_pass_filter(pcmf32, freq_thold, sample_rate); + } + + float energy_all = 0.0f; + float energy_last = 0.0f; + + for (size_t i = 0; i < n_samples; i++) { + energy_all += fabsf(pcmf32[i]); + + if (i >= n_samples - n_samples_last) { + energy_last += fabsf(pcmf32[i]); + } + } + + energy_all /= n_samples; + energy_last /= n_samples_last; + + if (verbose) { + fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold); + } + + if (energy_last > vad_thold*energy_all) { + return false; + } + + return true; +} + int main(int argc, char ** argv) { whisper_params params; @@ -173,33 +419,46 @@ int main(int argc, char ** argv) { return 1; } + params.keep_ms = std::min(params.keep_ms, params.step_ms); // cannot be more than step_ms + + const int n_samples_step = (params.step_ms *1e-3)*WHISPER_SAMPLE_RATE; + const int n_samples_len = (params.length_ms*1e-3)*WHISPER_SAMPLE_RATE; + const int n_samples_keep = (params.keep_ms *1e-3)*WHISPER_SAMPLE_RATE; + const int n_samples_30s = (30000 *1e-3)*WHISPER_SAMPLE_RATE; + + const int n_new_line = params.length_ms / params.step_ms - 1; // number of steps to print new line + + const bool use_vad = n_samples_step <= 0; + + params.no_timestamps = !use_vad; + params.no_context = use_vad; + params.max_tokens = 0; + // init audio - if (!audio_sdl_init(params.capture_id)) { - fprintf(stderr, "%s: audio_sdl_init() failed!\n", __func__); + audio_async audio(params.length_ms); + if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) { + fprintf(stderr, "%s: audio.init() failed!\n", __func__); return 1; } + audio.resume(); + + // whisper init + 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 = whisper_init(params.model.c_str()); - const int n_samples = (params.step_ms/1000.0)*WHISPER_SAMPLE_RATE; - const int n_samples_len = (params.length_ms/1000.0)*WHISPER_SAMPLE_RATE; - const int n_samples_30s = 30*WHISPER_SAMPLE_RATE; - const int n_samples_keep = 0.2*WHISPER_SAMPLE_RATE; - - std::vector pcmf32(n_samples_30s, 0.0f); - std::vector pcmf32_old; + std::vector pcmf32 (n_samples_30s, 0.0f); + std::vector pcmf32_old(n_samples_30s, 0.0f); + std::vector pcmf32_new(n_samples_30s, 0.0f); std::vector prompt_tokens; - const int n_new_line = params.length_ms / params.step_ms - 1; // print some info about the processing { @@ -211,23 +470,28 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__); } } - fprintf(stderr, "%s: processing %d samples (step = %.1f sec / len = %.1f sec), %d threads, lang = %s, task = %s, timestamps = %d ...\n", + fprintf(stderr, "%s: processing %d samples (step = %.1f sec / len = %.1f sec / keep = %.1f sec), %d threads, lang = %s, task = %s, timestamps = %d ...\n", __func__, - n_samples, - float(n_samples)/WHISPER_SAMPLE_RATE, - float(n_samples_len)/WHISPER_SAMPLE_RATE, + n_samples_step, + float(n_samples_step)/WHISPER_SAMPLE_RATE, + float(n_samples_len )/WHISPER_SAMPLE_RATE, + float(n_samples_keep)/WHISPER_SAMPLE_RATE, params.n_threads, params.language.c_str(), params.translate ? "translate" : "transcribe", params.no_timestamps ? 0 : 1); - fprintf(stderr, "%s: n_new_line = %d\n", __func__, n_new_line); + if (!use_vad) { + fprintf(stderr, "%s: n_new_line = %d\n", __func__, n_new_line); + } else { + fprintf(stderr, "%s: using VAD, will transcribe on speech activity\n", __func__); + } + fprintf(stderr, "\n"); } - SDL_PauseAudioDevice(g_dev_id_in, 0); - int n_iter = 0; + bool is_running = true; std::ofstream fout; @@ -242,6 +506,9 @@ int main(int argc, char ** argv) { printf("[Start speaking]"); fflush(stdout); + auto t_last = std::chrono::high_resolution_clock::now(); + const auto t_start = t_last; + // main audio loop while (is_running) { // handle Ctrl + C @@ -268,34 +535,63 @@ int main(int argc, char ** argv) { } // process new audio - if (n_iter > 0 && SDL_GetQueuedAudioSize(g_dev_id_in) > 2*n_samples*sizeof(float)) { - fprintf(stderr, "\n\n%s: WARNING: cannot process audio fast enough, dropping audio ...\n\n", __func__); - SDL_ClearQueuedAudio(g_dev_id_in); - } - while (SDL_GetQueuedAudioSize(g_dev_id_in) < n_samples*sizeof(float)) { - SDL_Delay(1); - } + if (!use_vad) { + while (true) { + audio.get(params.step_ms, pcmf32_new); - const int n_samples_new = SDL_GetQueuedAudioSize(g_dev_id_in)/sizeof(float); + if ((int) pcmf32_new.size() > 2*n_samples_step) { + fprintf(stderr, "\n\n%s: WARNING: cannot process audio fast enough, dropping audio ...\n\n", __func__); + audio.clear(); + continue; + } - // take one second from previous iteration - //const int n_samples_take = std::min((int) pcmf32_old.size(), std::max(0, n_samples_30s/30 - n_samples_new)); + if ((int) pcmf32_new.size() >= n_samples_step) { + audio.clear(); + break; + } - // take up to params.length_ms audio from previous iteration - const int n_samples_take = std::min((int) pcmf32_old.size(), std::max(0, n_samples_keep + n_samples_len - n_samples_new)); + SDL_Delay(1); + } - //printf("processing: take = %d, new = %d, old = %d\n", n_samples_take, n_samples_new, (int) pcmf32_old.size()); + const int n_samples_new = pcmf32_new.size(); - pcmf32.resize(n_samples_new + n_samples_take); + // take up to params.length_ms audio from previous iteration + const int n_samples_take = std::min((int) pcmf32_old.size(), std::max(0, n_samples_keep + n_samples_len - n_samples_new)); - for (int i = 0; i < n_samples_take; i++) { - pcmf32[i] = pcmf32_old[pcmf32_old.size() - n_samples_take + i]; - } + //printf("processing: take = %d, new = %d, old = %d\n", n_samples_take, n_samples_new, (int) pcmf32_old.size()); + + pcmf32.resize(n_samples_new + n_samples_take); + + for (int i = 0; i < n_samples_take; i++) { + pcmf32[i] = pcmf32_old[pcmf32_old.size() - n_samples_take + i]; + } + + memcpy(pcmf32.data() + n_samples_take, pcmf32_new.data(), n_samples_new*sizeof(float)); + + pcmf32_old = pcmf32; + } else { + const auto t_now = std::chrono::high_resolution_clock::now(); + const auto t_diff = std::chrono::duration_cast(t_now - t_last).count(); + + if (t_diff < 2000) { + std::this_thread::sleep_for(std::chrono::milliseconds(100)); + + continue; + } - SDL_DequeueAudio(g_dev_id_in, pcmf32.data() + n_samples_take, n_samples_new*sizeof(float)); + audio.get(2000, pcmf32_new); - pcmf32_old = pcmf32; + if (vad_simple(pcmf32_new, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, false)) { + audio.get(params.length_ms, pcmf32); + } else { + std::this_thread::sleep_for(std::chrono::milliseconds(100)); + + continue; + } + + t_last = t_now; + } // run the inference { @@ -307,7 +603,7 @@ int main(int argc, char ** argv) { wparams.print_timestamps = !params.no_timestamps; wparams.translate = params.translate; wparams.no_context = true; - wparams.single_segment = true; + wparams.single_segment = !use_vad; wparams.max_tokens = params.max_tokens; wparams.language = params.language.c_str(); wparams.n_threads = params.n_threads; @@ -325,12 +621,21 @@ int main(int argc, char ** argv) { // print result; { - printf("\33[2K\r"); + if (!use_vad) { + printf("\33[2K\r"); - // print long empty line to clear the previous line - printf("%s", std::string(100, ' ').c_str()); + // print long empty line to clear the previous line + printf("%s", std::string(100, ' ').c_str()); - printf("\33[2K\r"); + printf("\33[2K\r"); + } else { + const int64_t t1 = (t_last - t_start).count()/1000000; + const int64_t t0 = std::max(0.0, t1 - pcmf32.size()*1000.0/WHISPER_SAMPLE_RATE); + + printf("\n"); + printf("### Transcription %d START | t0 = %lld ms | t1 = %lld ms\n", n_iter, t0, t1); + printf("\n"); + } const int n_segments = whisper_full_n_segments(ctx); for (int i = 0; i < n_segments; ++i) { @@ -358,11 +663,16 @@ int main(int argc, char ** argv) { if (params.fname_out.length() > 0) { fout << std::endl; } + + if (use_vad){ + printf("\n"); + printf("### Transcription %d END\n", n_iter); + } } ++n_iter; - if ((n_iter % n_new_line) == 0) { + if (!use_vad && (n_iter % n_new_line) == 0) { printf("\n"); // keep part of the audio for next iteration to try to mitigate word boundary issues @@ -384,9 +694,7 @@ int main(int argc, char ** argv) { } } - if (g_dev_id_in >= 0) { - SDL_CloseAudioDevice(g_dev_id_in); - } + audio.pause(); whisper_print_timings(ctx); whisper_free(ctx);