| """Measure the frame offset of an alignment and print the correction to apply. |
| |
| Run per corpus, never assumed: English peaked at +0, msa-v7 at +1 under the same code. An |
| uncorrected one-frame offset is a milder version of the DTW failure that tripled WER. |
| """ |
| import json, sys |
| import numpy as np, soundfile as sf |
| HOP = 600 |
| cfg = json.load(open(sys.argv[2])) |
| inv = {v: k for k, v in cfg["vocab"].items()} |
| VOWELS = set("aeiouɑɐɒæɔəɚɛɜɪʊʌɤøœɨɯ") |
| QUIET = set(" ,.;:!?،؛؟") |
| rows = [json.loads(l) for l in open(sys.argv[1], encoding="utf-8")][:900] |
|
|
|
|
| def gap(shift): |
| v, q = [], [] |
| for r in rows: |
| n = r["frames"] |
| y, _ = sf.read(r["wav"], dtype="float32", start=int(r.get("start", 0)) * HOP, |
| frames=n * HOP) |
| if y.ndim > 1: y = y.mean(1) |
| if len(y) < n * HOP: y = np.pad(y, (0, n * HOP - len(y))) |
| e = 20 * np.log10(np.sqrt((y.reshape(n, HOP) ** 2).mean(1) + 1e-12)) |
| pos = 0 |
| for pid, d in zip(r["ids"], r["dur"]): |
| ch = inv.get(pid, "") |
| seg = e[max(pos + shift, 0): max(pos + d + shift, 0)] |
| if len(seg): |
| if ch in VOWELS: v.append(seg.mean()) |
| elif ch in QUIET: q.append(seg.mean()) |
| pos += d |
| return (np.mean(v) - np.mean(q)) if v and q else float("nan") |
|
|
|
|
| gs = {s: gap(s) for s in (-3, -2, -1, 0, 1, 2, 3)} |
| best = max(gs, key=lambda k: gs[k]) |
| print("sweep " + " ".join(f"{s:+d}:{g:5.2f}" for s, g in gs.items()), file=sys.stderr) |
| print(best) |
|
|
