| |
| """Regression test for H3 Audio Smear, the inverse of H3 Audio Recover. |
| |
| python tests/test_audio_smear.py |
| |
| Synthetic, no GPU, no files. Smear stretches a world-clock track onto the |
| dilated clock the smeared video init lives on; Recover compresses it back. |
| The pair has to close: |
| |
| 1. smeared length is sample-exact against sum(holds) frames |
| 2. smear -> recover returns the source length sample-exact |
| 3. the round trip preserves TIMING (envelope correlation), which is the |
| only property the seed has to carry - a phase vocoder does not |
| preserve phase, so sample-wise correlation is expected to be poor and |
| is not asserted |
| 4. an all-ones hold map is a no-op |
| |
| Why the timing matters: the seeded init exists so pass 2 renders a slowed |
| performance instead of inventing one at natural rate. If the stretch |
| misplaces events, the lips follow the wrong clock and Exact Recover / |
| Audio Recover compress something that was never slow. |
| |
| Exit code 0 = pass. |
| """ |
| import json |
| import math |
| import os |
| import sys |
|
|
| import torch |
|
|
| sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) |
| from motion import H3AudioRecover, H3AudioSmear |
|
|
| SR, FPS = 32000, 24 |
| FAIL = [] |
|
|
|
|
| def check(ok, msg): |
| print(("PASS " if ok else "FAIL ") + msg) |
| if not ok: |
| FAIL.append(msg) |
|
|
|
|
| def tone(n_frames): |
| """Amplitude-modulated tone: timing is legible in the envelope.""" |
| t = torch.arange(int(n_frames / FPS * SR)) / SR |
| x = torch.sin(2 * math.pi * 220 * t) * (0.5 + 0.5 * torch.sin(2 * math.pi * 5 * t)) |
| return {"waveform": x.float()[None, None, :], "sample_rate": SR} |
|
|
|
|
| def envelope(v, w=256): |
| return torch.nn.functional.avg_pool1d(v.abs()[None, None], w, w)[0, 0] |
|
|
|
|
| def corr(a, b): |
| n = min(a.shape[0], b.shape[0]) |
| a, b = a[:n], b[:n] |
| return float(torch.dot(a - a.mean(), b - b.mean()) |
| / (a.std() * b.std() * n + 1e-9)) |
|
|
|
|
| def main(): |
| |
| holds = [1] * 30 + [4] * 30 + [1] * 30 |
| hm = json.dumps({"holds": holds}) |
| src = tone(len(holds)) |
| n_src = src["waveform"].shape[-1] |
|
|
| sm = H3AudioSmear().smear(src, hm, FPS)[0] |
| want = int(round(sum(holds) * SR / FPS)) |
| check(abs(sm["waveform"].shape[-1] - want) <= 1, |
| f"smeared length sample-exact: {sm['waveform'].shape[-1]} vs {want}") |
|
|
| rt = H3AudioRecover().recover(sm, hm, FPS)[0] |
| check(abs(rt["waveform"].shape[-1] - n_src) <= 1, |
| f"round trip returns the world clock: {rt['waveform'].shape[-1]} vs {n_src}") |
|
|
| c = corr(envelope(src["waveform"].reshape(-1)), envelope(rt["waveform"].reshape(-1))) |
| check(c > 0.9, f"round trip preserves timing: envelope corr {c:+.3f} (> 0.9)") |
|
|
| |
| flat = json.dumps({"holds": [1] * 60}) |
| f = H3AudioSmear().smear(tone(60), flat, FPS)[0] |
| check(abs(f["waveform"].shape[-1] - int(round(60 * SR / FPS))) <= 1, |
| "all-ones hold map is a no-op on length") |
|
|
| print("\n" + ("ALL PASS" if not FAIL else f"{len(FAIL)} FAILED")) |
| return 1 if FAIL else 0 |
|
|
|
|
| if __name__ == "__main__": |
| sys.exit(main()) |
|
|
