whisper.cpp/examples/talk/talk.cpp

// Talk with AI
//

#include "whisper.h"
#include "gpt-2.h"

#include <SDL.h>
#include <SDL_audio.h>

#include <cassert>
#include <cstdio>
#include <fstream>
#include <mutex>
#include <regex>
#include <string>
#include <thread>
#include <vector>
#include <regex>

// 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;

    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;

    std::string person    = "Santa";
    std::string language  = "en";
    std::string model_wsp = "models/ggml-base.en.bin";
    std::string model_gpt = "models/ggml-gpt-2-117M.bin";
    std::string speak     = "./examples/talk/speak.sh";
    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 == "-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 == "-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 == "-mg"  || arg == "--model-gpt")     { params.model_gpt     = argv[++i]; }
        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
        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, "  -mg FILE, --model-gpt     [%-7s] gpt model file\n",                              params.model_gpt.c_str());
    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
    fprintf(stderr, "\n");
}

//
// SDL Audio capture
//

class audio_async {
public:
    audio_async(int len_ms);
    ~audio_async();

    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<float> & 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<float> m_audio;
    std::vector<float> 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 false;
    }

    SDL_SetHintWithPriority(SDL_HINT_AUDIO_RESAMPLING_MODE, "medium", SDL_HINT_OVERRIDE);

    {
        int nDevices = SDL_GetNumAudioDevices(SDL_TRUE);
        fprintf(stderr, "%s: found %d capture devices:\n", __func__, nDevices);
        for (int i = 0; i < nDevices; i++) {
            fprintf(stderr, "%s:    - Capture device #%d: '%s'\n", __func__, i, SDL_GetAudioDeviceName(i, SDL_TRUE));
        }
    }

    SDL_AudioSpec capture_spec_requested;
    SDL_AudioSpec capture_spec_obtained;

    SDL_zero(capture_spec_requested);
    SDL_zero(capture_spec_obtained);

    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));
        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__);
        m_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
    }

    if (!m_dev_id_in) {
        fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError());
        m_dev_id_in = 0;

        return false;
    } else {
        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);
        fprintf(stderr, "\n");
    }

    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<std::mutex> 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<std::mutex> 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<float> & 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<std::mutex> 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));
        }
    }
}

///////////////////////////

std::string trim(const std::string & s) {
    std::regex e("^\\s+|\\s+$");
    return std::regex_replace(s, e, "");
}

std::string replace(const std::string & s, const std::string & from, const std::string & to) {
    std::string result = s;
    size_t pos = 0;
    while ((pos = result.find(from, pos)) != std::string::npos) {
        result.replace(pos, from.length(), to);
        pos += to.length();
    }
    return result;
}

void high_pass_filter(std::vector<float> & 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<float> & 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 (int 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;
}

std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
    const auto t_start = std::chrono::high_resolution_clock::now();

    prob = 0.0f;
    t_ms = 0;

    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);

    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.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 =
R"(This is a dialogue between {0} (A) and a person (B). The dialogue so far is:

B: Hello {0}, how are you?
A: I'm fine, thank you.
{1}
Here is how {0} (A) continues the dialogue:

A:)";

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());

    // gpt init

    struct gpt2_context * ctx_gpt = gpt2_init(params.model_gpt.c_str());

    // 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;

    std::vector<float> pcmf32_cur;
    std::vector<float> pcmf32_prompt;

    gpt2_set_prompt(ctx_gpt, "");

    const int voice_id = rand()%6;

    fprintf(stderr, "gpt-2: prompt:\n");
    fprintf(stderr, "========================\n\n");
    fprintf(stderr, "%s\n", ::replace(k_prompt, "{0}", params.person).c_str());
    fprintf(stderr, "========================\n\n");

    // main loop
    while (is_running) {
        // handle Ctrl + C
        {
            SDL_Event event;
            while (SDL_PollEvent(&event)) {
                switch (event.type) {
                    case SDL_QUIT:
                        {
                            is_running = false;
                        } break;
                    default:
                        break;
                }
            }

            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, 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<gpt_vocab::id> tokens = gpt2_tokenize(ctx_gpt, text_heard.c_str());

                if (text_heard.empty() || tokens.empty() || force_speak) {
                    fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
                    audio.clear();

                    continue;
                }

                force_speak = false;

                fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", text_heard.c_str(), "\033[0m", (int) t_ms);

                std::string prompt_base = gpt2_get_prompt(ctx_gpt);

                std::string text_to_speak;

                {
                    prompt_base += "B: " + text_heard + "\n";

                    std::string prompt = ::replace(::replace(k_prompt, "{0}", params.person), "{1}", prompt_base);

                    text_to_speak = gpt2_gen_text(ctx_gpt, prompt.c_str(), params.max_tokens);
                    text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
                    text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of('\n'));

                    // remove first 2 lines of base prompt
                    if (n_iter > 4) {
                        {
                            const size_t pos = prompt_base.find_first_of('\n');
                            if (pos != std::string::npos) {
                                prompt_base = prompt_base.substr(pos + 1);
                            }
                        }
                        {
                            const size_t pos = prompt_base.find_first_of('\n');
                            if (pos != std::string::npos) {
                                prompt_base = prompt_base.substr(pos + 1);
                            }
                        }
                    }

                    prompt_base += "A:" + text_to_speak + "\n";

                    {
                        prompt = ::replace(::replace(k_prompt, "{0}", params.person), "{1}", prompt_base);

                        printf("===============\n");
                        printf("prompt:\n");
                        printf("%s\n", prompt.c_str());
                        printf("===============\n");
                    }
                }

                //printf("========================\n");
                //printf("gpt-2: prompt_base:\n%s\n", prompt_base.c_str());
                //printf("========================\n");

                gpt2_set_prompt(ctx_gpt, prompt_base.c_str());

                text_to_speak = ::replace(text_to_speak, params.person + ": ", "");
                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);

    return 0;
}