/* Command line front end: reads a wav, prints speech segments. * * Default output is the offline segmentation, which is what the upstream * get_speech_timestamps produces. --stream instead prints events at the moment * the streaming segmenter decides them, which is what a live pipeline sees. */ #include #include #include #include #include #include "svad.h" #include "wav.h" static double now_sec(void) { struct timespec ts; clock_gettime(CLOCK_MONOTONIC, &ts); return (double)ts.tv_sec + (double)ts.tv_nsec * 1e-9; } /* ru_maxrss is bytes on Darwin and kilobytes elsewhere. */ static double max_rss_mib(void) { struct rusage ru; getrusage(RUSAGE_SELF, &ru); #if defined(__APPLE__) return (double)ru.ru_maxrss / (1024.0 * 1024.0); #else return (double)ru.ru_maxrss / 1024.0; #endif } static void usage(FILE *f) { fputs("usage: svad [--stream] [--probs FILE] [--json] [--quiet] " "[--threshold X] [--neg-threshold X] [--min-speech-ms N] " "[--max-speech-s X] [--min-silence-ms N] [--speech-pad-ms N] " "[--weights FILE] [--scalar] [--dense-stft] [--no-ftz] INPUT.wav\n", f); } static int arg_int(int argc, char **argv, int *i, const char *name, long *out) { if (++*i >= argc) { fprintf(stderr, "svad: %s needs a value\n", name); return -1; } char *end; *out = strtol(argv[*i], &end, 10); if (*end) { fprintf(stderr, "svad: %s: bad integer '%s'\n", name, argv[*i]); return -1; } return 0; } static int arg_dbl(int argc, char **argv, int *i, const char *name, double *out) { if (++*i >= argc) { fprintf(stderr, "svad: %s needs a value\n", name); return -1; } char *end; *out = strtod(argv[*i], &end); if (*end) { fprintf(stderr, "svad: %s: bad number '%s'\n", name, argv[*i]); return -1; } return 0; } int main(int argc, char **argv) { const char *wav_path = NULL, *probs_path = NULL, *weights_path = NULL; int stream = 0, json = 0, quiet = 0; unsigned flags = 0; svad_segment_params sp; svad_segment_params_default(&sp); long lv; double dv; for (int i = 1; i < argc; ++i) { const char *a = argv[i]; if (strcmp(a, "-h") == 0 || strcmp(a, "--help") == 0) { usage(stdout); return 0; } else if (strcmp(a, "--stream") == 0) { stream = 1; } else if (strcmp(a, "--json") == 0) { json = 1; } else if (strcmp(a, "--quiet") == 0) { quiet = 1; } else if (strcmp(a, "--scalar") == 0) { flags |= SVAD_SCALAR_KERNELS; } else if (strcmp(a, "--dense-stft") == 0) { flags |= SVAD_DENSE_STFT; } else if (strcmp(a, "--no-ftz") == 0) { flags |= SVAD_NO_FTZ; } else if (strcmp(a, "--probs") == 0 && i + 1 < argc) { probs_path = argv[++i]; } else if (strcmp(a, "--weights") == 0 && i + 1 < argc) { weights_path = argv[++i]; } else if (strcmp(a, "--threshold") == 0) { if (arg_dbl(argc, argv, &i, a, &dv)) return 2; sp.threshold = (float)dv; } else if (strcmp(a, "--neg-threshold") == 0) { if (arg_dbl(argc, argv, &i, a, &dv)) return 2; sp.neg_threshold = (float)dv; } else if (strcmp(a, "--max-speech-s") == 0) { if (arg_dbl(argc, argv, &i, a, &dv)) return 2; sp.max_speech_s = dv; } else if (strcmp(a, "--min-speech-ms") == 0) { if (arg_int(argc, argv, &i, a, &lv)) return 2; sp.min_speech_ms = (int)lv; } else if (strcmp(a, "--min-silence-ms") == 0) { if (arg_int(argc, argv, &i, a, &lv)) return 2; sp.min_silence_ms = (int)lv; } else if (strcmp(a, "--speech-pad-ms") == 0) { if (arg_int(argc, argv, &i, a, &lv)) return 2; sp.speech_pad_ms = (int)lv; } else if (a[0] == '-' && a[1]) { fprintf(stderr, "svad: unknown option '%s'\n", a); usage(stderr); return 2; } else if (!wav_path) { wav_path = a; } else { fprintf(stderr, "svad: unexpected argument '%s'\n", a); return 2; } } if (!wav_path) { usage(stderr); return 2; } wav_t wav; const char *err = NULL; if (wav_read_mono_s16(wav_path, &wav, &err) != 0) { fprintf(stderr, "svad: %s: %s\n", wav_path, err ? err : "read failed"); return 1; } if (wav.sample_rate != 8000 && wav.sample_rate != 16000) { fprintf(stderr, "svad: %s: %d Hz; only 8000 and 16000 are supported\n", wav_path, wav.sample_rate); wav_free(&wav); return 1; } const double t_init0 = now_sec(); svad_t *v = svad_create_ex(weights_path, wav.sample_rate, flags); const double init_ms = (now_sec() - t_init0) * 1e3; if (!v) { fprintf(stderr, "svad: initialization failed%s\n", weights_path ? " (bad weight file?)" : ""); wav_free(&wav); return 1; } /* A trailing partial chunk is zero-padded and processed, matching the * reference implementation's chunking of the same file. */ const int chunk = svad_chunk_samples(v); const size_t n_chunks = (wav.count + (size_t)chunk - 1) / (size_t)chunk; float *probs = malloc((n_chunks ? n_chunks : 1) * sizeof(float)); if (!probs) { fprintf(stderr, "svad: out of memory\n"); svad_destroy(v); wav_free(&wav); return 1; } svad_segmenter seg; svad_segmenter_init(&seg, wav.sample_rate); seg.threshold = sp.threshold; if (sp.neg_threshold > 0.0f) seg.neg_threshold = sp.neg_threshold; else seg.neg_threshold = sp.threshold - 0.15f; seg.min_silence_samples = (int64_t)wav.sample_rate * sp.min_silence_ms / 1000; seg.speech_pad_samples = (int64_t)wav.sample_rate * sp.speech_pad_ms / 1000; int16_t tail[512] = {0}; const double t0 = now_sec(); for (size_t i = 0; i < n_chunks; ++i) { const size_t off = i * (size_t)chunk; const int16_t *src = wav.samples + off; if (off + (size_t)chunk > wav.count) { memset(tail, 0, sizeof tail); memcpy(tail, src, (wav.count - off) * sizeof(int16_t)); src = tail; } probs[i] = svad_process_s16(v, src); if (stream) { const svad_event ev = svad_segmenter_update(&seg, probs[i]); if (ev.kind == SVAD_EVENT_SPEECH_START) printf("start %.3f\n", (double)ev.sample / wav.sample_rate); else if (ev.kind == SVAD_EVENT_SPEECH_END) printf("end %.3f\n", (double)ev.sample / wav.sample_rate); } } const double elapsed = now_sec() - t0; if (probs_path) { FILE *pf = fopen(probs_path, "wb"); if (!pf || fwrite(probs, sizeof(float), n_chunks, pf) != n_chunks) { fprintf(stderr, "svad: cannot write %s\n", probs_path); if (pf) fclose(pf); free(probs); svad_destroy(v); wav_free(&wav); return 1; } fclose(pf); } if (!stream) { const size_t n = svad_speech_timestamps(probs, n_chunks, (int64_t)wav.count, wav.sample_rate, &sp, NULL, 0); svad_segment *segs = n ? malloc(n * sizeof *segs) : NULL; if (n && !segs) { fprintf(stderr, "svad: out of memory\n"); free(probs); svad_destroy(v); wav_free(&wav); return 1; } svad_speech_timestamps(probs, n_chunks, (int64_t)wav.count, wav.sample_rate, &sp, segs, n); if (json) printf("["); for (size_t i = 0; i < n; ++i) { const double s = (double)segs[i].start / wav.sample_rate; const double e = (double)segs[i].end / wav.sample_rate; if (json) printf("%s{\"start\":%.3f,\"end\":%.3f,\"start_sample\":%lld,\"end_sample\":%lld}", i ? "," : "", s, e, (long long)segs[i].start, (long long)segs[i].end); else printf("%8.3f %8.3f %10lld %10lld\n", s, e, (long long)segs[i].start, (long long)segs[i].end); } if (json) printf("]\n"); free(segs); } if (!quiet) { const double audio_s = (double)wav.count / wav.sample_rate; fprintf(stderr, "%s: %.2f s audio, %zu chunks, %s backend, init %.3f ms, " "%.1f us/chunk, %.0fx realtime, max rss %.1f MiB\n", wav_path, audio_s, n_chunks, svad_backend_name(v), init_ms, n_chunks ? elapsed / (double)n_chunks * 1e6 : 0.0, elapsed > 0 ? audio_s / elapsed : 0.0, max_rss_mib()); } free(probs); svad_destroy(v); wav_free(&wav); return 0; }