/* Inputs the model will never see in training but will see in production. * * The model normalises its input by the window's own standard deviation, so * digital silence divides by roughly the epsilon in the normalisation and a * constant offset does the same. Those are the cases that produce infinities * if the epsilon is dropped or applied in the wrong place, and a whole window * of silence is entirely ordinary in a real recording. */ #include "pyseg.h" #include "test_util.h" #include #include #include static unsigned rng_state = 987654321u; static float next_float(void) { rng_state = rng_state * 1664525u + 1013904223u; return (float)((rng_state >> 8) & 0xFFFF) / 32768.0f - 1.0f; } static void check_output(const char *label, const float *probs) { for (int t = 0; t < PYSEG_FRAMES; ++t) { const float *row = probs + (size_t)t * PYSEG_CLASSES; double sum = 0.0; for (int i = 0; i < PYSEG_CLASSES; ++i) { if (!isfinite(row[i]) || row[i] < 0.0f || row[i] > 1.0f) { TEST_ASSERT(0, "%s: frame %d class %d is %g", label, t, i, row[i]); return; } sum += row[i]; } if (fabs(sum - 1.0) > 1e-5) { TEST_ASSERT(0, "%s: frame %d sums to %.9f", label, t, sum); return; } } TEST_ASSERT(1, "%s", label); } int main(void) { pyseg_t *v = pyseg_create(); TEST_ASSERT(v != NULL, "create failed"); if (!v) return TEST_REPORT(); float *wav = malloc(PYSEG_WINDOW_SAMPLES * sizeof(float)); float *probs = malloc((size_t)PYSEG_FRAMES * PYSEG_CLASSES * sizeof(float)); float *active = malloc((size_t)PYSEG_FRAMES * PYSEG_SPEAKERS * sizeof(float)); int16_t *pcm = malloc(PYSEG_WINDOW_SAMPLES * sizeof(int16_t)); memset(wav, 0, PYSEG_WINDOW_SAMPLES * sizeof(float)); pyseg_process_f32(v, wav, probs); check_output("digital silence", probs); for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) wav[i] = 0.5f; pyseg_process_f32(v, wav, probs); check_output("constant offset", probs); for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) wav[i] = (i & 1) ? 1.0f : -1.0f; pyseg_process_f32(v, wav, probs); check_output("full-scale square wave", probs); for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) wav[i] = next_float() * 4.0f; pyseg_process_f32(v, wav, probs); check_output("out-of-range noise", probs); for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) wav[i] = next_float() * 1e-8f; pyseg_process_f32(v, wav, probs); check_output("near-denormal noise", probs); /* A single impulse in an otherwise silent window: the normalisation * amplifies the silence around it by the full dynamic range. */ memset(wav, 0, PYSEG_WINDOW_SAMPLES * sizeof(float)); wav[PYSEG_WINDOW_SAMPLES / 2] = 1.0f; pyseg_process_f32(v, wav, probs); check_output("lone impulse", probs); for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) pcm[i] = (int16_t)(i & 1 ? 32767 : -32768); pyseg_process_s16(v, pcm, probs); check_output("s16 full scale", probs); /* The s16 entry point must be the f32 one with a fixed scaling. */ for (int i = 0; i < PYSEG_WINDOW_SAMPLES; ++i) { pcm[i] = (int16_t)((int)(next_float() * 16000.0f)); wav[i] = (float)pcm[i] * (1.0f / 32768.0f); } float *from_s16 = malloc((size_t)PYSEG_FRAMES * PYSEG_CLASSES * sizeof(float)); pyseg_process_s16(v, pcm, from_s16); pyseg_process_f32(v, wav, probs); TEST_ASSERT(memcmp(from_s16, probs, (size_t)PYSEG_FRAMES * PYSEG_CLASSES * sizeof(float)) == 0, "s16 and f32 entry points disagree on identical audio"); /* Decoding must land inside the powerset and stay consistent with it. */ pyseg_to_multilabel(probs, active); for (int t = 0; t < PYSEG_FRAMES; ++t) { const float *a = active + (size_t)t * PYSEG_SPEAKERS; int count = 0; for (int s = 0; s < PYSEG_SPEAKERS; ++s) { if (a[s] != 0.0f && a[s] != 1.0f) { TEST_ASSERT(0, "frame %d speaker %d is %g, not a hard decision", t, s, a[s]); break; } count += a[s] != 0.0f; } if (count > 2) { TEST_ASSERT(0, "frame %d has %d speakers, more than the powerset allows", t, count); break; } } TEST_ASSERT(1, "multilabel decoding stays inside the powerset"); TEST_ASSERT(pyseg_powerset_class(-1, -1) == 0, "empty set is not class 0"); TEST_ASSERT(pyseg_powerset_class(0, -1) == 1, "speaker 0 alone is not class 1"); TEST_ASSERT(pyseg_powerset_class(2, -1) == 3, "speaker 2 alone is not class 3"); TEST_ASSERT(pyseg_powerset_class(0, 1) == 4, "speakers 0 and 1 are not class 4"); TEST_ASSERT(pyseg_powerset_class(1, 2) == 6, "speakers 1 and 2 are not class 6"); free(from_s16); free(pcm); free(active); free(probs); free(wav); pyseg_destroy(v); pyseg_destroy(NULL); return TEST_REPORT(); }