whispercpp 1.3.6 → 1.3.8

data/ext/sources/examples/server/server.cpp

@@ -56,11 +56,11 @@ inline void signal_handler(int signal) {
 
 struct server_params
 {
-    std::string hostname = "127.0.0.1";
-    std::string public_path = "examples/server/public";
-    std::string request_path = "";
+    std::string hostname       = "127.0.0.1";
+    std::string public_path    = "examples/server/public";
+    std::string request_path   = "";
     std::string inference_path = "/inference";
-    std::string tmp_dir = ".";
+    std::string tmp_dir        = ".";
 
     int32_t port          = 8080;
     int32_t read_timeout  = 600;
@@ -87,50 +87,47 @@ struct whisper_params {
     float logprob_thold   = -1.00f;
     float temperature     =  0.00f;
     float temperature_inc =  0.20f;
-    float no_speech_thold = 0.6f;
-
-    bool debug_mode      = false;
-    bool translate       = false;
-    bool detect_language = false;
-    bool diarize         = false;
-    bool tinydiarize     = false;
-    bool split_on_word   = false;
-    bool no_fallback     = false;
-    bool print_special   = false;
-    bool print_colors    = false;
-    bool print_realtime  = false;
-    bool print_progress  = false;
-    bool no_timestamps   = false;
-    bool use_gpu         = true;
-    bool flash_attn      = true;
-    int32_t gpu_device   = 0;
-    bool suppress_nst    = false;
-    bool no_context      = true;
+    float no_speech_thold =  0.6f;
+
+    bool debug_mode                = false;
+    bool translate                 = false;
+    bool detect_language           = false;
+    bool diarize                   = false;
+    bool tinydiarize               = false;
+    bool split_on_word             = false;
+    bool no_fallback               = false;
+    bool print_special             = false;
+    bool print_colors              = false;
+    bool print_realtime            = false;
+    bool print_progress            = false;
+    bool no_timestamps             = false;
+    bool token_timestamps          = true;
+    bool use_gpu                   = true;
+    bool flash_attn                = true;
+    int32_t gpu_device             = 0;
+    bool suppress_nst              = false;
+    bool no_context                = true;
     bool no_language_probabilities = false;
-
-    std::string language        = "en";
-    std::string prompt          = "";
-    std::string font_path       = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
-    std::string model           = "models/ggml-base.en.bin";
-
-    std::string response_format     = json_format;
-
-    // [TDRZ] speaker turn string
-    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line
-
+    bool carry_initial_prompt      = false;
+
+    std::string language               = "en";
+    std::string prompt                 = "";
+    std::string font_path              = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    std::string model                  = "models/ggml-base.en.bin";
+    std::string response_format        = json_format;
+    std::string tdrz_speaker_turn      = " [SPEAKER_TURN]"; // TODO: set from command line
     std::string openvino_encode_device = "CPU";
-
-    std::string dtw = "";
+    std::string dtw                    = "";
 
     // Voice Activity Detection (VAD) parameters
-    bool        vad           = false;
-    std::string vad_model     = "";
-    float       vad_threshold = 0.5f;
-    int         vad_min_speech_duration_ms = 250;
+    bool        vad                         = false;
+    std::string vad_model                   = "";
+    float       vad_threshold               = 0.5f;
+    int         vad_min_speech_duration_ms  = 250;
     int         vad_min_silence_duration_ms = 100;
-    float       vad_max_speech_duration_s = FLT_MAX;
-    int         vad_speech_pad_ms = 30;
-    float       vad_samples_overlap = 0.1f;
+    float       vad_max_speech_duration_s   = FLT_MAX;
+    int         vad_speech_pad_ms           = 30;
+    float       vad_samples_overlap         = 0.1f;
 };
 
 void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params, const server_params& sparams) {
@@ -138,51 +135,52 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     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, "  -p N,      --processors N      [%-7d] number of processors to use during computation\n", params.n_processors);
-    fprintf(stderr, "  -ot N,     --offset-t N        [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
-    fprintf(stderr, "  -on N,     --offset-n N        [%-7d] segment index offset\n",                           params.offset_n);
-    fprintf(stderr, "  -d  N,     --duration N        [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
-    fprintf(stderr, "  -mc N,     --max-context N     [%-7d] maximum number of text context tokens to store\n", params.max_context);
-    fprintf(stderr, "  -ml N,     --max-len N         [%-7d] maximum segment length in characters\n",           params.max_len);
-    fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
-    fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
-    fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
-    fprintf(stderr, "  -ac N,     --audio-ctx N       [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
-    fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
-    fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
-    fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
-    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
-    fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
-    fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
-    fprintf(stderr, "  -tdrz,     --tinydiarize       [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
-    fprintf(stderr, "  -nf,       --no-fallback       [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
-    fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
-    fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
-    fprintf(stderr, "  -pr,       --print-realtime    [%-7s] print output in realtime\n",                       params.print_realtime ? "true" : "false");
-    fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
-    fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
-    fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
-    fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
-    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
-    fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
-    fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
+    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, "  -p N,      --processors N              [%-7d] number of processors to use during computation\n", params.n_processors);
+    fprintf(stderr, "  -ot N,     --offset-t N                [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
+    fprintf(stderr, "  -on N,     --offset-n N                [%-7d] segment index offset\n",                           params.offset_n);
+    fprintf(stderr, "  -d  N,     --duration N                [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
+    fprintf(stderr, "  -mc N,     --max-context N             [%-7d] maximum number of text context tokens to store\n", params.max_context);
+    fprintf(stderr, "  -ml N,     --max-len N                 [%-7d] maximum segment length in characters\n",           params.max_len);
+    fprintf(stderr, "  -sow,      --split-on-word             [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
+    fprintf(stderr, "  -bo N,     --best-of N                 [%-7d] number of best candidates to keep\n",              params.best_of);
+    fprintf(stderr, "  -bs N,     --beam-size N               [%-7d] beam size for beam search\n",                      params.beam_size);
+    fprintf(stderr, "  -ac N,     --audio-ctx N               [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
+    fprintf(stderr, "  -wt N,     --word-thold N              [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
+    fprintf(stderr, "  -et N,     --entropy-thold N           [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
+    fprintf(stderr, "  -lpt N,    --logprob-thold N           [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
+    fprintf(stderr, "  -debug,    --debug-mode                [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
+    fprintf(stderr, "  -tr,       --translate                 [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
+    fprintf(stderr, "  -di,       --diarize                   [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
+    fprintf(stderr, "  -tdrz,     --tinydiarize               [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
+    fprintf(stderr, "  -nf,       --no-fallback               [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
+    fprintf(stderr, "  -ps,       --print-special             [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pc,       --print-colors              [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
+    fprintf(stderr, "  -pr,       --print-realtime            [%-7s] print output in realtime\n",                       params.print_realtime ? "true" : "false");
+    fprintf(stderr, "  -pp,       --print-progress            [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
+    fprintf(stderr, "  -nt,       --no-timestamps             [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
+    fprintf(stderr, "  -l LANG,   --language LANG             [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
+    fprintf(stderr, "  -dl,       --detect-language           [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
+    fprintf(stderr, "             --prompt PROMPT             [%-7s] initial prompt\n",                                 params.prompt.c_str());
+    fprintf(stderr, "             --carry-initial-prompt      [%-7s] always prepend initial prompt\n",                  params.carry_initial_prompt ? "true" : "false");
+    fprintf(stderr, "  -m FNAME,  --model FNAME               [%-7s] model path\n",                                     params.model.c_str());
+    fprintf(stderr, "  -oved D,   --ov-e-device DNAME         [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
     // server params
-    fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n", params.dtw.c_str());
-    fprintf(stderr, "  --host HOST,                   [%-7s] Hostname/ip-adress for the server\n", sparams.hostname.c_str());
-    fprintf(stderr, "  --port PORT,                   [%-7d] Port number for the server\n", sparams.port);
-    fprintf(stderr, "  --public PATH,                 [%-7s] Path to the public folder\n", sparams.public_path.c_str());
-    fprintf(stderr, "  --request-path PATH,           [%-7s] Request path for all requests\n", sparams.request_path.c_str());
-    fprintf(stderr, "  --inference-path PATH,         [%-7s] Inference path for all requests\n", sparams.inference_path.c_str());
-    fprintf(stderr, "  --convert,                     [%-7s] Convert audio to WAV, requires ffmpeg on the server\n", sparams.ffmpeg_converter ? "true" : "false");
-    fprintf(stderr, "  --tmp-dir,                     [%-7s] Temporary directory for ffmpeg transcoded files\n", sparams.tmp_dir.c_str());
-    fprintf(stderr, "  -sns,      --suppress-nst      [%-7s] suppress non-speech tokens\n", params.suppress_nst ? "true" : "false");
-    fprintf(stderr, "  -nth N,    --no-speech-thold N [%-7.2f] no speech threshold\n",   params.no_speech_thold);
-    fprintf(stderr, "  -ng,       --no-gpu            [%-7s] do not use gpu\n", params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -dev N,    --device N          [%-7d] GPU device ID (default: 0)\n", params.gpu_device);
-    fprintf(stderr, "  -fa,       --flash-attn        [%-7s] enable flash attention\n", params.flash_attn ? "true" : "false");
-    fprintf(stderr, "  -nfa,      --no-flash-attn     [%-7s] disable flash attention\n", params.flash_attn ? "false" : "true");
+    fprintf(stderr, "  -dtw MODEL --dtw MODEL                 [%-7s] compute token-level timestamps\n",                          params.dtw.c_str());
+    fprintf(stderr, "  --host HOST,                           [%-7s] Hostname/ip-adress for the server\n",                       sparams.hostname.c_str());
+    fprintf(stderr, "  --port PORT,                           [%-7d] Port number for the server\n",                              sparams.port);
+    fprintf(stderr, "  --public PATH,                         [%-7s] Path to the public folder\n",                               sparams.public_path.c_str());
+    fprintf(stderr, "  --request-path PATH,                   [%-7s] Request path for all requests\n",                           sparams.request_path.c_str());
+    fprintf(stderr, "  --inference-path PATH,                 [%-7s] Inference path for all requests\n",                         sparams.inference_path.c_str());
+    fprintf(stderr, "  --convert,                             [%-7s] Convert audio to WAV, requires ffmpeg on the server\n",     sparams.ffmpeg_converter ? "true" : "false");
+    fprintf(stderr, "  --tmp-dir,                             [%-7s] Temporary directory for ffmpeg transcoded files\n",         sparams.tmp_dir.c_str());
+    fprintf(stderr, "  -sns,      --suppress-nst              [%-7s] suppress non-speech tokens\n",                              params.suppress_nst ? "true" : "false");
+    fprintf(stderr, "  -nth N,    --no-speech-thold N         [%-7.2f] no speech threshold\n",                                   params.no_speech_thold);
+    fprintf(stderr, "  -ng,       --no-gpu                    [%-7s] do not use gpu\n",                                          params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -dev N,    --device N                  [%-7d] GPU device ID (default: 0)\n",                              params.gpu_device);
+    fprintf(stderr, "  -fa,       --flash-attn                [%-7s] enable flash attention\n",                                  params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -nfa,      --no-flash-attn             [%-7s] disable flash attention\n",                                 params.flash_attn ? "false" : "true");
     fprintf(stderr, "  -nlp,      --no-language-probabilities [%-7s] exclude language probabilities from verbose_json output\n", params.no_language_probabilities ? "true" : "false");
     // Voice Activity Detection (VAD) parameters
     fprintf(stderr, "\nVoice Activity Detection (VAD) options:\n");
@@ -190,10 +188,8 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -vm FNAME, --vad-model FNAME               [%-7s] VAD model path\n",                                   params.vad_model.c_str());
     fprintf(stderr, "  -vt N,     --vad-threshold N               [%-7.2f] VAD threshold for speech recognition\n",           params.vad_threshold);
     fprintf(stderr, "  -vspd N,   --vad-min-speech-duration-ms  N [%-7d] VAD min speech duration (0.0-1.0)\n",                params.vad_min_speech_duration_ms);
-    fprintf(stderr, "  -vsd N,    --vad-min-silence-duration-ms N [%-7d] VAD min silence duration (to split segments)\n",      params.vad_min_silence_duration_ms);
-    fprintf(stderr, "  -vmsd N,   --vad-max-speech-duration-s   N [%-7s] VAD max speech duration (auto-split longer)\n",      params.vad_max_speech_duration_s == FLT_MAX ?
-                                                                                                                                  std::string("FLT_MAX").c_str() :
-                                                                                                                                  std::to_string(params.vad_max_speech_duration_s).c_str());
+    fprintf(stderr, "  -vsd N,    --vad-min-silence-duration-ms N [%-7d] VAD min silence duration (to split segments)\n",     params.vad_min_silence_duration_ms);
+    fprintf(stderr, "  -vmsd N,   --vad-max-speech-duration-s   N [%-7s] VAD max speech duration (auto-split longer)\n",      params.vad_max_speech_duration_s == FLT_MAX ? std::string("FLT_MAX").c_str() : std::to_string(params.vad_max_speech_duration_s).c_str());
     fprintf(stderr, "  -vp N,     --vad-speech-pad-ms           N [%-7d] VAD speech padding (extend segments)\n",             params.vad_speech_pad_ms);
     fprintf(stderr, "  -vo N,     --vad-samples-overlap         N [%-7.2f] VAD samples overlap (seconds between segments)\n", params.vad_samples_overlap);
     fprintf(stderr, "\n");
@@ -211,63 +207,64 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
             whisper_print_usage(argc, argv, params, sparams);
             exit(0);
         }
-        else if (arg == "-t"    || arg == "--threads")         { params.n_threads       = std::stoi(argv[++i]); }
-        else if (arg == "-p"    || arg == "--processors")      { params.n_processors    = std::stoi(argv[++i]); }
-        else if (arg == "-ot"   || arg == "--offset-t")        { params.offset_t_ms     = std::stoi(argv[++i]); }
-        else if (arg == "-on"   || arg == "--offset-n")        { params.offset_n        = std::stoi(argv[++i]); }
-        else if (arg == "-d"    || arg == "--duration")        { params.duration_ms     = std::stoi(argv[++i]); }
-        else if (arg == "-mc"   || arg == "--max-context")     { params.max_context     = std::stoi(argv[++i]); }
-        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
-        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
-        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
-        else if (arg == "-ac"   || arg == "--audio-ctx")       { params.audio_ctx       = std::stoi(argv[++i]); }
-        else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
-        else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
-        else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
-        else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
-        else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
-        else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
-        else if (arg == "-tdrz" || arg == "--tinydiarize")     { params.tinydiarize     = true; }
-        else if (arg == "-sow"  || arg == "--split-on-word")   { params.split_on_word   = true; }
-        else if (arg == "-nf"   || arg == "--no-fallback")     { params.no_fallback     = true; }
-        else if (arg == "-fp"   || arg == "--font-path")       { params.font_path       = argv[++i]; }
-        else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
-        else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
-        else if (arg == "-pr"   || arg == "--print-realtime")  { params.print_realtime  = true; }
-        else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
-        else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
-        else if (arg == "-l"    || arg == "--language")        { params.language        = argv[++i]; }
-        else if (arg == "-dl"   || arg == "--detect-language") { params.detect_language = true; }
-        else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
-        else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
-        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
-        else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
-        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
-        else if (arg == "-dev"  || arg == "--device")          { params.gpu_device      = std::stoi(argv[++i]); }
-        else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
-        else if (arg == "-nfa"  || arg == "--no-flash-attn")   { params.flash_attn      = false; }
-        else if (arg == "-sns"  || arg == "--suppress-nst")    { params.suppress_nst    = true; }
-        else if (arg == "-nth"  || arg == "--no-speech-thold") { params.no_speech_thold = std::stof(argv[++i]); }
-        else if (arg == "-nlp"  || arg == "--no-language-probabilities") { params.no_language_probabilities = true; }
+        else if (arg == "-t"     || arg == "--threads")                   { params.n_threads                 = std::stoi(argv[++i]); }
+        else if (arg == "-p"     || arg == "--processors")                { params.n_processors              = std::stoi(argv[++i]); }
+        else if (arg == "-ot"    || arg == "--offset-t")                  { params.offset_t_ms               = std::stoi(argv[++i]); }
+        else if (arg == "-on"    || arg == "--offset-n")                  { params.offset_n                  = std::stoi(argv[++i]); }
+        else if (arg == "-d"     || arg == "--duration")                  { params.duration_ms               = std::stoi(argv[++i]); }
+        else if (arg == "-mc"    || arg == "--max-context")               { params.max_context               = std::stoi(argv[++i]); }
+        else if (arg == "-ml"    || arg == "--max-len")                   { params.max_len                   = std::stoi(argv[++i]); }
+        else if (arg == "-bo"    || arg == "--best-of")                   { params.best_of                   = std::stoi(argv[++i]); }
+        else if (arg == "-bs"    || arg == "--beam-size")                 { params.beam_size                 = std::stoi(argv[++i]); }
+        else if (arg == "-ac"    || arg == "--audio-ctx")                 { params.audio_ctx                 = std::stoi(argv[++i]); }
+        else if (arg == "-wt"    || arg == "--word-thold")                { params.word_thold                = std::stof(argv[++i]); }
+        else if (arg == "-et"    || arg == "--entropy-thold")             { params.entropy_thold             = std::stof(argv[++i]); }
+        else if (arg == "-lpt"   || arg == "--logprob-thold")             { params.logprob_thold             = std::stof(argv[++i]); }
+        else if (arg == "-debug" || arg == "--debug-mode")                { params.debug_mode                = true; }
+        else if (arg == "-tr"    || arg == "--translate")                 { params.translate                 = true; }
+        else if (arg == "-di"    || arg == "--diarize")                   { params.diarize                   = true; }
+        else if (arg == "-tdrz"  || arg == "--tinydiarize")               { params.tinydiarize               = true; }
+        else if (arg == "-sow"   || arg == "--split-on-word")             { params.split_on_word             = true; }
+        else if (arg == "-nf"    || arg == "--no-fallback")               { params.no_fallback               = true; }
+        else if (arg == "-fp"    || arg == "--font-path")                 { params.font_path                 = argv[++i]; }
+        else if (arg == "-ps"    || arg == "--print-special")             { params.print_special             = true; }
+        else if (arg == "-pc"    || arg == "--print-colors")              { params.print_colors              = true; }
+        else if (arg == "-pr"    || arg == "--print-realtime")            { params.print_realtime            = true; }
+        else if (arg == "-pp"    || arg == "--print-progress")            { params.print_progress            = true; }
+        else if (arg == "-nt"    || arg == "--no-timestamps")             { params.no_timestamps             = true; }
+        else if (arg == "-l"     || arg == "--language")                  { params.language                  = argv[++i]; }
+        else if (arg == "-dl"    || arg == "--detect-language")           { params.detect_language           = true; }
+        else if (                   arg == "--prompt")                    { params.prompt                    = argv[++i]; }
+        else if (                   arg == "--carry-initial-prompt")      { params.carry_initial_prompt      = true; }
+        else if (arg == "-m"     || arg == "--model")                     { params.model                     = argv[++i]; }
+        else if (arg == "-oved"  || arg == "--ov-e-device")               { params.openvino_encode_device    = argv[++i]; }
+        else if (arg == "-dtw"   || arg == "--dtw")                       { params.dtw                       = argv[++i]; }
+        else if (arg == "-ng"    || arg == "--no-gpu")                    { params.use_gpu                   = false; }
+        else if (arg == "-dev"   || arg == "--device")                    { params.gpu_device                = std::stoi(argv[++i]); }
+        else if (arg == "-fa"    || arg == "--flash-attn")                { params.flash_attn                = true; }
+        else if (arg == "-nfa"   || arg == "--no-flash-attn")             { params.flash_attn                = false; }
+        else if (arg == "-sns"   || arg == "--suppress-nst")              { params.suppress_nst              = true; }
+        else if (arg == "-nth"   || arg == "--no-speech-thold")           { params.no_speech_thold           = std::stof(argv[++i]); }
+        else if (arg == "-nlp"   || arg == "--no-language-probabilities") { params.no_language_probabilities = true; }
 
         // server params
-        else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
-        else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
-        else if (                  arg == "--public")          { sparams.public_path = argv[++i]; }
-        else if (                  arg == "--request-path")    { sparams.request_path = argv[++i]; }
-        else if (                  arg == "--inference-path")  { sparams.inference_path = argv[++i]; }
-        else if (                  arg == "--convert")         { sparams.ffmpeg_converter     = true; }
-        else if (                  arg == "--tmp-dir")         { sparams.tmp_dir     = argv[++i]; }
+        else if (                   arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
+        else if (                   arg == "--host")            { sparams.hostname    = argv[++i]; }
+        else if (                   arg == "--public")          { sparams.public_path = argv[++i]; }
+        else if (                   arg == "--request-path")    { sparams.request_path = argv[++i]; }
+        else if (                   arg == "--inference-path")  { sparams.inference_path = argv[++i]; }
+        else if (                   arg == "--convert")         { sparams.ffmpeg_converter     = true; }
+        else if (                   arg == "--tmp-dir")         { sparams.tmp_dir     = argv[++i]; }
 
         // Voice Activity Detection (VAD)
-        else if (                  arg == "--vad")                         { params.vad                         = true; }
-        else if (arg == "-vm"   || arg == "--vad-model")                   { params.vad_model                   = argv[++i]; }
-        else if (arg == "-vt"   || arg == "--vad-threshold")               { params.vad_threshold               = std::stof(argv[++i]); }
-        else if (arg == "-vspd" || arg == "--vad-min-speech-duration-ms")  { params.vad_min_speech_duration_ms  = std::stoi(argv[++i]); }
-        else if (arg == "-vsd"  || arg == "--vad-min-silence-duration-ms") { params.vad_min_silence_duration_ms = std::stoi(argv[++i]); }
-        else if (arg == "-vmsd" || arg == "--vad-max-speech-duration-s")   { params.vad_max_speech_duration_s   = std::stof(argv[++i]); }
-        else if (arg == "-vp"   || arg == "--vad-speech-pad-ms")           { params.vad_speech_pad_ms           = std::stoi(argv[++i]); }
-        else if (arg == "-vo"   || arg == "--vad-samples-overlap")         { params.vad_samples_overlap         = std::stof(argv[++i]); }
+        else if (                   arg == "--vad")                         { params.vad                         = true; }
+        else if (arg == "-vm"    || arg == "--vad-model")                   { params.vad_model                   = argv[++i]; }
+        else if (arg == "-vt"    || arg == "--vad-threshold")               { params.vad_threshold               = std::stof(argv[++i]); }
+        else if (arg == "-vspd"  || arg == "--vad-min-speech-duration-ms")  { params.vad_min_speech_duration_ms  = std::stoi(argv[++i]); }
+        else if (arg == "-vsd"   || arg == "--vad-min-silence-duration-ms") { params.vad_min_silence_duration_ms = std::stoi(argv[++i]); }
+        else if (arg == "-vmsd"  || arg == "--vad-max-speech-duration-s")   { params.vad_max_speech_duration_s   = std::stof(argv[++i]); }
+        else if (arg == "-vp"    || arg == "--vad-speech-pad-ms")           { params.vad_speech_pad_ms           = std::stoi(argv[++i]); }
+        else if (arg == "-vo"    || arg == "--vad-samples-overlap")         { params.vad_samples_overlap         = std::stof(argv[++i]); }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params, sparams);
@@ -318,10 +315,10 @@ std::string generate_temp_filename(const std::string &path, const std::string &p
     return ss.str();
 }
 
-bool convert_to_wav(const std::string & temp_filename, std::string & error_resp) {
+bool convert_to_wav(const std::string & temp_filename, std::string & error_resp, bool stereo) {
     std::ostringstream cmd_stream;
     std::string converted_filename_temp = temp_filename + "_temp.wav";
-    cmd_stream << "ffmpeg -i \"" << temp_filename << "\" -y -ar 16000 -ac 1 -c:a pcm_s16le \"" << converted_filename_temp << "\" 2>&1";
+    cmd_stream << "ffmpeg -i \"" << temp_filename << "\" -y -ar 16000 -ac " << (stereo ? 2 : 1) << " -c:a pcm_s16le \"" << converted_filename_temp << "\" 2>&1";
     std::string cmd = cmd_stream.str();
 
     int status = std::system(cmd.c_str());
@@ -344,7 +341,7 @@ bool convert_to_wav(const std::string & temp_filename, std::string & error_resp)
     return true;
 }
 
-std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
+std::string estimate_diarization_speaker(const std::vector<std::vector<float>> & pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
     std::string speaker = "";
     const int64_t n_samples = pcmf32s[0].size();
 
@@ -454,7 +451,7 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
     }
 }
 
-std::string output_str(struct whisper_context * ctx, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+std::string output_str(struct whisper_context * ctx, const whisper_params & params, const std::vector<std::vector<float>> & pcmf32s) {
     std::stringstream result;
     const int n_segments = whisper_full_n_segments(ctx);
     for (int i = 0; i < n_segments; ++i) {
@@ -526,6 +523,10 @@ void get_req_parameters(const Request & req, whisper_params & params)
     {
         params.logprob_thold = std::stof(req.get_file_value("logprob_thold").content);
     }
+    if (req.has_file("no_speech_thold"))
+    {
+        params.no_speech_thold = std::stof(req.get_file_value("no_speech_thold").content);
+    }
     if (req.has_file("debug_mode"))
     {
         params.debug_mode = parse_str_to_bool(req.get_file_value("debug_mode").content);
@@ -550,6 +551,12 @@ void get_req_parameters(const Request & req, whisper_params & params)
     {
         params.no_timestamps = parse_str_to_bool(req.get_file_value("no_timestamps").content);
     }
+    if (req.has_file("token_timestamps"))
+    {
+        params.token_timestamps = parse_str_to_bool(req.get_file_value("token_timestamps").content);
+    } else {
+        params.token_timestamps = !params.no_timestamps;
+    }
     if (req.has_file("language"))
     {
         params.language = req.get_file_value("language").content;
@@ -562,6 +569,10 @@ void get_req_parameters(const Request & req, whisper_params & params)
     {
         params.prompt = req.get_file_value("prompt").content;
     }
+    if (req.has_file("carry_initial_prompt"))
+    {
+        params.carry_initial_prompt = parse_str_to_bool(req.get_file_value("carry_initial_prompt").content);
+    }
     if (req.has_file("response_format"))
     {
         params.response_format = req.get_file_value("response_format").content;
@@ -690,10 +701,10 @@ int main(int argc, char ** argv) {
         if (params.dtw == "large.v3") {
             cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
         }
-        if (params.dtw == "large.v3.turbo") { 
+        if (params.dtw == "large.v3.turbo") {
             cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3_TURBO;
         }
-        
+
         if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
             fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
             return 3;
@@ -755,6 +766,7 @@ int main(int argc, char ** argv) {
     -F file="@&lt;file-path&gt;" \
     -F temperature="0.0" \
     -F temperature_inc="0.2" \
+    -F no_speech_thold="0.6" \
     -F response_format="json"
         </pre>
 
@@ -817,7 +829,7 @@ int main(int argc, char ** argv) {
         }
         auto audio_file = req.get_file_value("file");
 
-        // check non-required fields
+        whisper_params params = default_params;
         get_req_parameters(req, params);
 
         std::string filename{audio_file.filename};
@@ -836,7 +848,7 @@ int main(int argc, char ** argv) {
             temp_file.close();
 
             std::string error_resp = "{\"error\":\"Failed to execute ffmpeg command.\"}";
-            const bool is_converted = convert_to_wav(temp_filename, error_resp);
+            const bool is_converted = convert_to_wav(temp_filename, error_resp, params.diarize);
             if (!is_converted) {
                 res.status = 500;
                 res.set_content(error_resp, "application/json");
@@ -856,8 +868,7 @@ int main(int argc, char ** argv) {
             // remove temp file
             std::remove(temp_filename.c_str());
         } else {
-            if (!::read_audio_data(audio_file.content, pcmf32, pcmf32s, params.diarize))
-            {
+            if (!::read_audio_data(audio_file.content.data(), audio_file.content.size(), pcmf32, pcmf32s, params.diarize)) {
                 fprintf(stderr, "error: failed to read audio data\n");
                 const std::string error_resp = "{\"error\":\"failed to read audio data\"}";
                 res.status = 400;
@@ -928,18 +939,19 @@ int main(int argc, char ** argv) {
             wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
 
             wparams.initial_prompt   = params.prompt.c_str();
+            wparams.carry_initial_prompt = params.carry_initial_prompt;
 
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
 
             wparams.temperature      = params.temperature;
-            wparams.no_speech_thold = params.no_speech_thold;
+            wparams.no_speech_thold  = params.no_speech_thold;
             wparams.temperature_inc  = params.temperature_inc;
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
 
             wparams.no_timestamps    = params.no_timestamps;
-            wparams.token_timestamps = !params.no_timestamps;
+            wparams.token_timestamps = params.token_timestamps;
             wparams.no_context       = params.no_context;
 
             wparams.suppress_nst     = params.suppress_nst;
@@ -1043,7 +1055,7 @@ int main(int argc, char ** argv) {
             res.set_content(ss.str(), "text/vtt");
         } else if (params.response_format == vjson_format) {
             /* try to match openai/whisper's Python format */
-            std::string results = output_str(ctx, params, pcmf32s); 
+            std::string results = output_str(ctx, params, pcmf32s);
             json jres = json{
                 {"task", params.translate ? "translate" : "transcribe"},
                 {"language", whisper_lang_str_full(whisper_full_lang_id(ctx))},
@@ -1078,6 +1090,14 @@ int main(int argc, char ** argv) {
                     segment["end"] = whisper_full_get_segment_t1(ctx, i) * 0.01;
                 }
 
+                if (params.diarize && pcmf32s.size() == 2) {
+                    segment["speaker"] = estimate_diarization_speaker(
+                        pcmf32s,
+                        whisper_full_get_segment_t0(ctx, i),
+                        whisper_full_get_segment_t1(ctx, i),
+                        true);
+                }
+
                 float total_logprob = 0;
                 const int n_tokens = whisper_full_n_tokens(ctx, i);
                 for (int j = 0; j < n_tokens; ++j) {
@@ -1087,10 +1107,29 @@ int main(int argc, char ** argv) {
                     }
 
                     segment["tokens"].push_back(token.id);
-                    json word = json{{"word", whisper_full_get_token_text(ctx, i, j)}};
-                    if (!params.no_timestamps) {
+                    std::string word_text = whisper_full_get_token_text(ctx, i, j);
+                    int64_t word_t1 = token.t1;
+
+                    while (j + 1 < n_tokens && utf8_trailing_bytes_needed(word_text) > 0) {
+                        const whisper_token_data next_token = whisper_full_get_token_data(ctx, i, j + 1);
+                        // Keep verbose_json tokens free of EOT ids, matching the pre-merge server behavior.
+                        if (next_token.id >= whisper_token_eot(ctx)) {
+                            break;
+                        }
+
+                        ++j;
+                        segment["tokens"].push_back(next_token.id);
+                        word_text += whisper_full_get_token_text(ctx, i, j);
+                        if (next_token.t1 > -1) {
+                            word_t1 = next_token.t1;
+                        }
+                        total_logprob += next_token.plog;
+                    }
+
+                    json word = json{{"word", word_text}};
+                    if (!params.no_timestamps && params.token_timestamps) {
                         word["start"] = token.t0 * 0.01;
-                        word["end"] = token.t1 * 0.01;
+                        word["end"] = word_t1 * 0.01;
                         word["t_dtw"] = token.t_dtw;
                     }
                     word["probability"] = token.p;
@@ -1120,9 +1159,6 @@ int main(int argc, char ** argv) {
             res.set_content(jres.dump(-1, ' ', false, json::error_handler_t::replace),
                             "application/json");
         }
-
-        // reset params to their defaults
-        params = default_params;
     });
     svr->Post(sparams.request_path + "/load", [&](const Request &req, Response &res){
         std::lock_guard<std::mutex> lock(whisper_mutex);

data/ext/sources/examples/vad-speech-segments/speech.cpp

@@ -26,6 +26,7 @@ static void vad_print_usage(int /*argc*/, char ** argv, const cli_params & param
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options] file\n", argv[0]);
     fprintf(stderr, "supported audio formats: flac, mp3, ogg, wav\n");
+    fprintf(stderr, "Note: Output timestamps are in centiseconds (1/100th of a second).\n");
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h,        --help                          [default] show this help message and exit\n");
@@ -64,7 +65,7 @@ static bool vad_params_parse(int argc, char ** argv, cli_params & params) {
         else if (arg == "-ug"   || arg == "--use-gpu")                     { params.use_gpu                     = true; }
         else if (arg == "-vm"   || arg == "--vad-model")                   { params.vad_model                   = ARGV_NEXT; }
         else if (arg == "-vt"   || arg == "--vad-threshold")               { params.vad_threshold               = std::stof(ARGV_NEXT); }
-        else if (arg == "-vsd"  || arg == "--vad-min-speech-duration-ms")  { params.vad_min_speech_duration_ms  = std::stoi(ARGV_NEXT); }
+        else if (arg == "-vspd" || arg == "--vad-min-speech-duration-ms")  { params.vad_min_speech_duration_ms  = std::stoi(ARGV_NEXT); }
         else if (arg == "-vsd"  || arg == "--vad-min-silence-duration-ms") { params.vad_min_speech_duration_ms  = std::stoi(ARGV_NEXT); }
         else if (arg == "-vmsd" || arg == "--vad-max-speech-duration-s")   { params.vad_max_speech_duration_s   = std::stof(ARGV_NEXT); }
         else if (arg == "-vp"   || arg == "--vad-speech-pad-ms")           { params.vad_speech_pad_ms           = std::stoi(ARGV_NEXT); }
@@ -122,7 +123,7 @@ int main(int argc, char ** argv) {
         return 3;
     }
 
-    // Get the the vad segements using the probabilities that have been computed
+    // Get the the vad segments using the probabilities that have been computed
     // previously and stored in the whisper_vad_context.
     struct whisper_vad_params params = whisper_vad_default_params();
     params.threshold = cli_params.vad_threshold;