whispercpp 1.3.5 → 1.3.7

data/ext/sources/examples/cli/cli.cpp

@@ -77,6 +77,7 @@ struct whisper_params {
     bool log_score       = false;
     bool use_gpu         = true;
     bool flash_attn      = true;
+    int32_t gpu_device   = 0;
     bool suppress_nst    = false;
     bool carry_initial_prompt = false;
 
@@ -129,6 +130,10 @@ static char * requires_value_error(const std::string & arg) {
 }
 
 static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    if (const char * env_device = std::getenv("WHISPER_ARG_DEVICE")) {
+        params.gpu_device = std::stoi(env_device);
+    }
+
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
@@ -146,6 +151,10 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
+        if (arg == "--version") {
+            fprintf(stdout, "whisper.cpp version: %s\n", whisper_version());
+            exit(0);
+        }
         #define ARGV_NEXT (((i + 1) < argc) ? argv[++i] : requires_value_error(arg))
         else if (arg == "-t"    || arg == "--threads")              { params.n_threads       = std::stoi(ARGV_NEXT); }
         else if (arg == "-p"    || arg == "--processors")           { params.n_processors    = std::stoi(ARGV_NEXT); }
@@ -195,6 +204,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
         else if (arg == "-dtw"  || arg == "--dtw")                  { params.dtw             = ARGV_NEXT; }
         else if (arg == "-ls"   || arg == "--log-score")            { params.log_score       = true; }
         else if (arg == "-ng"   || arg == "--no-gpu")               { params.use_gpu         = false; }
+        else if (arg == "-dev"  || arg == "--device")               { params.gpu_device      = std::stoi(ARGV_NEXT); }
         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; }
@@ -228,6 +238,7 @@ static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h,        --help                 [default] show this help message and exit\n");
+    fprintf(stderr, "             --version              show version information 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);
@@ -276,6 +287,7 @@ static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params
     fprintf(stderr, "  -dtw MODEL --dtw MODEL            [%-7s] compute token-level timestamps\n",                 params.dtw.c_str());
     fprintf(stderr, "  -ls,       --log-score            [%-7s] log best decoder scores of tokens\n",              params.log_score?"true":"false");
     fprintf(stderr, "  -ng,       --no-gpu               [%-7s] disable 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, "  -sns,      --suppress-nst         [%-7s] suppress non-speech tokens\n",                     params.suppress_nst ? "true" : "false");
@@ -731,18 +743,47 @@ static void output_json(
                     if (full) {
                         start_arr("tokens");
                         const int n = whisper_full_n_tokens(ctx, i);
-                        for (int j = 0; j < n; ++j) {
-                            auto token = whisper_full_get_token_data(ctx, i, j);
+
+                        // Merge adjacent tokens whose bytes together form a
+                        // single UTF-8 codepoint. Multi-byte characters (CJK
+                        // in particular) can end up split across whisper
+                        // tokens, which used to produce invalid UTF-8 in the
+                        // JSON string. Refs issue #1798.
+                        struct merged_token {
+                            std::string        text;
+                            whisper_token_data data;
+                            int64_t            t1;
+                        };
+                        std::vector<merged_token> merged;
+                        merged.reserve(n);
+                        for (int j = 0; j < n; ) {
+                            auto tok = whisper_full_get_token_data(ctx, i, j);
+                            merged_token m{ whisper_token_to_str(ctx, tok.id), tok, tok.t1 };
+                            ++j;
+                            while (j < n && utf8_trailing_bytes_needed(m.text) > 0) {
+                                auto tok_next = whisper_full_get_token_data(ctx, i, j);
+                                m.text += whisper_token_to_str(ctx, tok_next.id);
+                                if (tok_next.t1 > -1) {
+                                    m.t1 = tok_next.t1;
+                                }
+                                ++j;
+                            }
+                            merged.push_back(std::move(m));
+                        }
+
+                        const int nm = (int) merged.size();
+                        for (int j = 0; j < nm; ++j) {
+                            const auto & mt = merged[j];
                             start_obj(nullptr);
-                                value_s("text", whisper_token_to_str(ctx, token.id), false);
-                                if(token.t0 > -1 && token.t1 > -1) {
+                                value_s("text", mt.text.c_str(), false);
+                                if (mt.data.t0 > -1 && mt.t1 > -1) {
                                     // If we have per-token timestamps, write them out
-                                    times_o(token.t0, token.t1, false);
+                                    times_o(mt.data.t0, mt.t1, false);
                                 }
-                                value_i("id", token.id, false);
-                                value_f("p", token.p, false);
-                                value_f("t_dtw", token.t_dtw, true);
-                            end_obj(j == (n - 1));
+                                value_i("id", mt.data.id, false);
+                                value_f("p", mt.data.p, false);
+                                value_f("t_dtw", mt.data.t_dtw, true);
+                            end_obj(j == (nm - 1));
                         }
                         end_arr(!params.diarize && !params.tinydiarize);
                     }
@@ -1003,6 +1044,7 @@ int main(int argc, char ** argv) {
     struct whisper_context_params cparams = whisper_context_default_params();
 
     cparams.use_gpu    = params.use_gpu;
+    cparams.gpu_device = params.gpu_device;
     cparams.flash_attn = params.flash_attn;
 
     if (!params.dtw.empty()) {

data/ext/sources/examples/common-ggml.cpp

@@ -73,6 +73,8 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_IQ1_M:
         case GGML_FTYPE_MOSTLY_BF16:
         case GGML_FTYPE_MOSTLY_MXFP4:
+        case GGML_FTYPE_MOSTLY_NVFP4:
+        case GGML_FTYPE_MOSTLY_Q1_0:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -213,6 +215,8 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_TQ1_0:
                 case GGML_TYPE_TQ2_0:
                 case GGML_TYPE_MXFP4:
+                case GGML_TYPE_NVFP4:
+                case GGML_TYPE_Q1_0:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

data/ext/sources/examples/common-whisper.cpp

@@ -34,91 +34,35 @@
 #include <cstring>
 #include <fstream>
 
-#ifdef WHISPER_FFMPEG
-// as implemented in ffmpeg_trancode.cpp only embedded in common lib if whisper built with ffmpeg support
-extern bool ffmpeg_decode_audio(const std::string & ifname, std::vector<uint8_t> & wav_data);
+#ifdef WHISPER_COMMON_FFMPEG
+// as implemented in ffmpeg-trancode.cpp only embedded in common lib if whisper built with ffmpeg support
+extern bool ffmpeg_decode_audio(const std::string & ifname, std::vector<uint8_t> & wav_data, int out_sample_rate = WHISPER_SAMPLE_RATE);
 #endif
 
-bool read_audio_data(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
-    std::vector<uint8_t> audio_data; // used for pipe input from stdin or ffmpeg decoding output
-
+// extract f32 PCM frames from an initialized decoder, downmix to mono and keep the stereo split
+static bool read_audio_from_decoder(ma_decoder & decoder, std::vector<float> & pcmf32, std::vector<std::vector<float>> & pcmf32s, bool stereo) {
     ma_result result;
-    ma_decoder_config decoder_config;
-    ma_decoder decoder;
-
-    decoder_config = ma_decoder_config_init(ma_format_f32, stereo ? 2 : 1, WHISPER_SAMPLE_RATE);
-
-    if (fname == "-") {
-		#ifdef _WIN32
-		_setmode(_fileno(stdin), _O_BINARY);
-		#endif
-
-		uint8_t buf[1024];
-		while (true)
-		{
-			const size_t n = fread(buf, 1, sizeof(buf), stdin);
-			if (n == 0) {
-				break;
-			}
-			audio_data.insert(audio_data.end(), buf, buf + n);
-		}
-
-		if ((result = ma_decoder_init_memory(audio_data.data(), audio_data.size(), &decoder_config, &decoder)) != MA_SUCCESS) {
-
-			fprintf(stderr, "Error: failed to open audio data from stdin (%s)\n", ma_result_description(result));
-
-			return false;
-		}
-
-		fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, audio_data.size());
-    }
-    else if (((result = ma_decoder_init_file(fname.c_str(), &decoder_config, &decoder)) != MA_SUCCESS)) {
-#if defined(WHISPER_FFMPEG)
-		if (ffmpeg_decode_audio(fname, audio_data) != 0) {
-			fprintf(stderr, "error: failed to ffmpeg decode '%s'\n", fname.c_str());
-
-			return false;
-		}
-
-		if ((result = ma_decoder_init_memory(audio_data.data(), audio_data.size(), &decoder_config, &decoder)) != MA_SUCCESS) {
-			fprintf(stderr, "error: failed to read audio data as wav (%s)\n", ma_result_description(result));
-
-			return false;
-		}
-#else
-		if ((result = ma_decoder_init_memory(fname.c_str(), fname.size(), &decoder_config, &decoder)) != MA_SUCCESS) {
-			fprintf(stderr, "error: failed to read audio data as wav (%s)\n", ma_result_description(result));
-
-			return false;
-		}
-#endif
-    }
-
     ma_uint64 frame_count;
     ma_uint64 frames_read;
 
     if ((result = ma_decoder_get_length_in_pcm_frames(&decoder, &frame_count)) != MA_SUCCESS) {
-		fprintf(stderr, "error: failed to retrieve the length of the audio data (%s)\n", ma_result_description(result));
-
-		return false;
+        fprintf(stderr, "error: failed to retrieve the length of the audio data (%s)\n", ma_result_description(result));
+        return false;
     }
 
     pcmf32.resize(stereo ? frame_count*2 : frame_count);
 
     if ((result = ma_decoder_read_pcm_frames(&decoder, pcmf32.data(), frame_count, &frames_read)) != MA_SUCCESS) {
-		fprintf(stderr, "error: failed to read the frames of the audio data (%s)\n", ma_result_description(result));
-
-		return false;
+        fprintf(stderr, "error: failed to read the frames of the audio data (%s)\n", ma_result_description(result));
+        return false;
     }
 
     if (stereo) {
         std::vector<float> stereo_data = pcmf32;
         pcmf32.resize(frame_count);
-
         for (uint64_t i = 0; i < frame_count; i++) {
             pcmf32[i] = (stereo_data[2*i] + stereo_data[2*i + 1]);
         }
-
         pcmf32s.resize(2);
         pcmf32s[0].resize(frame_count);
         pcmf32s[1].resize(frame_count);
@@ -128,11 +72,111 @@ bool read_audio_data(const std::string & fname, std::vector<float>& pcmf32, std:
         }
     }
 
-    ma_decoder_uninit(&decoder);
-
     return true;
 }
 
+bool read_audio_data(const std::string & fname, std::vector<float> & pcmf32, std::vector<std::vector<float>> & pcmf32s, bool stereo) {
+    std::vector<uint8_t> audio_data; // used for pipe input from stdin or ffmpeg decoding output
+
+    ma_result result;
+    ma_decoder_config decoder_config;
+
+    struct decoder_guard {
+        ma_decoder decoder;
+        bool initialized = false;
+        ma_decoder * operator&() { return &decoder; }
+        ~decoder_guard() {
+            if (initialized) {
+                ma_decoder_uninit(&decoder);
+            }
+        }
+    };
+    decoder_guard decoder{};
+
+    decoder_config = ma_decoder_config_init(ma_format_f32, stereo ? 2 : 1, WHISPER_SAMPLE_RATE);
+
+    if (fname == "-") {
+#ifdef _WIN32
+        _setmode(_fileno(stdin), _O_BINARY);
+#endif
+
+        uint8_t buf[1024];
+        while (true)
+        {
+            const size_t n = fread(buf, 1, sizeof(buf), stdin);
+            if (n == 0) {
+                break;
+            }
+            audio_data.insert(audio_data.end(), buf, buf + n);
+        }
+
+        result = ma_decoder_init_memory(audio_data.data(), audio_data.size(), &decoder_config, &decoder);
+        if (result != MA_SUCCESS) {
+            fprintf(stderr, "%s: failed to open audio data from stdin (%s)\n", __func__, ma_result_description(result));
+            return false;
+        }
+        decoder.initialized = true;
+
+        fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, audio_data.size());
+    } else {
+        fprintf(stderr, "%s: reading audio data from '%s' ...\n", __func__, fname.c_str());
+
+        // first try miniaudio. if it fails (or skipped) - try ffmpeg
+        {
+            const char * skip = getenv("WHISPER_COMMON_MINIAUDIO_SKIP");
+            if (!skip || strlen(skip) == 0 || strcmp(skip, "0") == 0) {
+                fprintf(stderr, "%s: trying to decode with miniaudio\n", __func__);
+
+                result = ma_decoder_init_file(fname.c_str(), &decoder_config, &decoder);
+                if (result == MA_SUCCESS) {
+                    decoder.initialized = true;
+                }
+            } else {
+                fprintf(stderr, "%s: skipping miniaudio\n", __func__);
+            }
+        }
+
+#if defined(WHISPER_COMMON_FFMPEG)
+        if (!decoder.initialized) {
+            fprintf(stderr, "%s: trying to decode with ffmpeg\n", __func__);
+
+            if (ffmpeg_decode_audio(fname, audio_data) != 0) {
+                fprintf(stderr, "%s: failed to ffmpeg decode\n", __func__);
+                return false;
+            }
+            result = ma_decoder_init_memory(audio_data.data(), audio_data.size(), &decoder_config, &decoder);
+            if (result != MA_SUCCESS) {
+                fprintf(stderr, "%s: failed to read audio data as wav (%s)\n", __func__, ma_result_description(result));
+                return false;
+            }
+            decoder.initialized = true;
+        }
+#endif
+
+        if (!decoder.initialized) {
+            fprintf(stderr, "%s: failed to read audio data\n", __func__);
+            return false;
+        }
+    }
+
+    return read_audio_from_decoder(decoder.decoder, pcmf32, pcmf32s, stereo);
+}
+
+// decode audio bytes already held in memory
+bool read_audio_data(const char * buffer, size_t buffer_size, std::vector<float> & pcmf32, std::vector<std::vector<float>> & pcmf32s, bool stereo) {
+    ma_decoder_config decoder_config = ma_decoder_config_init(ma_format_f32, stereo ? 2 : 1, WHISPER_SAMPLE_RATE);
+    ma_decoder decoder;
+
+    if (ma_decoder_init_memory(buffer, buffer_size, &decoder_config, &decoder) != MA_SUCCESS) {
+        fprintf(stderr, "error: failed to decode audio data from memory buffer\n");
+        return false;
+    }
+
+    bool ok = read_audio_from_decoder(decoder, pcmf32, pcmf32s, stereo);
+    ma_decoder_uninit(&decoder);
+    return ok;
+}
+
 //  500 -> 00:05.000
 // 6000 -> 01:00.000
 std::string to_timestamp(int64_t t, bool comma) {
@@ -154,6 +198,34 @@ int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate) {
     return std::max(0, std::min((int) n_samples - 1, (int) ((t*whisper_sample_rate)/100)));
 }
 
+int utf8_trailing_bytes_needed(const std::string & s) {
+    const int n = (int) s.size();
+    int i = n - 1;
+    while (i >= 0 && ((unsigned char) s[i] & 0xC0) == 0x80) {
+        --i;
+    }
+    if (i < 0) {
+        return 0;
+    }
+
+    const unsigned char c = (unsigned char) s[i];
+    int expected;
+    if ((c & 0x80) == 0x00) {
+        expected = 1;
+    } else if ((c & 0xE0) == 0xC0) {
+        expected = 2;
+    } else if ((c & 0xF0) == 0xE0) {
+        expected = 3;
+    } else if ((c & 0xF8) == 0xF0) {
+        expected = 4;
+    } else {
+        return 0;
+    }
+
+    const int have = n - i;
+    return have >= expected ? 0 : (expected - have);
+}
+
 bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id) {
     std::ofstream speak_file(path.c_str());
     if (speak_file.fail()) {

data/ext/sources/examples/common-whisper.h

@@ -14,11 +14,22 @@ bool read_audio_data(
         std::vector<std::vector<float>> & pcmf32s,
         bool stereo);
 
+// decode audio bytes already held in memory (uploaded file, network buffer)
+bool read_audio_data(
+        const char * buffer,
+        size_t buffer_size,
+        std::vector<float> & pcmf32,
+        std::vector<std::vector<float>> & pcmf32s,
+        bool stereo);
+
 // convert timestamp to string, 6000 -> 01:00.000
 std::string to_timestamp(int64_t t, bool comma = false);
 
 // given a timestamp get the sample
 int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate);
 
+// Returns the number of trailing bytes still needed for s to end on a complete UTF-8 codepoint.
+int utf8_trailing_bytes_needed(const std::string & s);
+
 // write text to file, and call system("command voice_id file")
 bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id);