canary-180m-flash empty-transcript reproduction (transcribe.cpp)

empty-crimson-falcon.wav

empty-golden-anchor.wav

GIST-README.md

canary-180m-flash empty-transcript reproduction

Companion files for an issue on handy-computer/transcribe.cpp.

file	canary-180m-flash	Apple Speech, same bytes
empty-golden-anchor.wav (7.5s)	0 chars, status 0, 0.05s	"Golden Anchor 99, recorder path guard verified."
empty-crimson-falcon.wav (7.3s)	0 chars, status 0, 0.04s	"Crimson Falcon 42. Single install verified."
ok-same-mic-control.wav (6s)	55 chars ✓	—

All 16 kHz mono. All machine-generated: macOS say played through the laptop speakers and recorded by the built-in mic. No human speech.

ok-same-mic-control.wav was recorded minutes later through the same mic, same room, same speaker-playback path, and transcribes correctly — so the failures are not a capture-quality problem.

probe.c links the xcframework and calls transcribe_run directly, with no application code:

clang -O2 -o probe probe.c -I <fw>/CTranscribe.framework/Headers \
    -F <fw> -framework CTranscribe -Wl,-rpath,<fw>
./probe canary-180m-flash-Q8_0.gguf empty-golden-anchor.wav empty-crimson-falcon.wav ok-same-mic-control.wav

ok-same-mic-control.wav

probe.c

#include <transcribe.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>
#include <time.h>

// Minimal 16-bit mono WAV reader: walk chunks to find "data".
static float *read_wav(const char *path, int *out_n) {
    FILE *f = fopen(path, "rb"); if (!f) return NULL;
    char id[4]; uint32_t sz;
    fseek(f, 12, SEEK_SET);                    // skip RIFF/WAVE
    while (fread(id, 1, 4, f) == 4 && fread(&sz, 4, 1, f) == 1) {
        if (!memcmp(id, "data", 4)) {
            int n = sz / 2;
            int16_t *pcm = malloc(sz);
            if (fread(pcm, 1, sz, f) != sz) { free(pcm); fclose(f); return NULL; }
            float *out = malloc(sizeof(float) * n);
            for (int i = 0; i < n; i++) out[i] = pcm[i] / 32768.0f;   // same as CanaryEngine.normalize
            free(pcm); fclose(f); *out_n = n; return out;
        }
        fseek(f, (sz + 1) & ~1u, SEEK_CUR);
    }
    fclose(f); return NULL;
}

int main(int argc, char **argv) {
    if (argc < 3) { fprintf(stderr, "usage: probe <model.gguf> <wav>...\n"); return 2; }
    struct transcribe_model_load_params mp;
    transcribe_model_load_params_init(&mp);    // device left unset = automatic (v0.2 change)
    struct transcribe_model *model = NULL;
    transcribe_status st = transcribe_model_load_file(argv[1], &mp, &model);
    if ((int)st != 0 || !model) { fprintf(stderr, "model load failed: %d\n", (int)st); return 1; }
    printf("loaded transcribe.cpp %s\n", TRANSCRIBE_VERSION);

    for (int a = 2; a < argc; a++) {
        int n = 0; float *pcm = read_wav(argv[a], &n);
        if (!pcm) { printf("%-34s  WAV READ FAILED\n", argv[a]); continue; }
        struct transcribe_session_params sp; transcribe_session_params_init(&sp);
        struct transcribe_session *sess = NULL;
        if ((int)transcribe_session_init(model, &sp, &sess) != 0 || !sess) {
            printf("%-34s  SESSION FAILED\n", argv[a]); free(pcm); continue;
        }
        struct transcribe_run_params rp; transcribe_run_params_init(&rp);
        rp.language = "en"; rp.target_language = "en";     // same as Warbler
        struct timespec t0, t1; clock_gettime(CLOCK_MONOTONIC, &t0);
        transcribe_status rs = transcribe_run(sess, pcm, n, &rp);
        clock_gettime(CLOCK_MONOTONIC, &t1);
        double dt = (t1.tv_sec - t0.tv_sec) + (t1.tv_nsec - t0.tv_nsec) / 1e9;
        const char *txt = ((int)rs == 0) ? transcribe_full_text(sess) : NULL;
        const char *base = strrchr(argv[a], '/'); base = base ? base + 1 : argv[a];
        printf("%-34s %5.2fs  status=%d  chars=%zu  %s\n", base, dt, (int)rs,
               txt ? strlen(txt) : 0,
               (txt && *txt) ? txt : "<<< EMPTY >>>");
        transcribe_session_free(sess); free(pcm);
    }
    transcribe_model_free(model);
    return 0;
}