whisper probe: 6 low-voicing candidates

kaggle/whisper_probe.py

@@ -0,0 +1,144 @@
+"""Which whisper clips actually work as Ming prompts?
+
+a01 was chosen by ear for its breathy delivery, but as a prompt it produces
+hallucinated speech ("字幕by索兰娅" for 「后来店关了，歌还在。」), and the teacher
+does this too, so it is not a distillation artefact. Six other low-voicing clips
+were mined; this bakes each and asks two questions:
+
+  1. does the clone say the right words (CER), and
+  2. does it stay breathy (voicing of the output against the prompt's own)
+
+A clip that speaks correctly but loses the whisper is a fail: the breathiness is
+the reason we want it.
+
+Outputs whisper_probe/{slot}_{i}.wav and metrics.json.
+"""
+import json
+import os
+import sys
+import time
+import traceback
+from pathlib import Path
+
+import numpy as np
+
+ROOT = Path(os.environ.get("WORK_ROOT", "/kaggle/working"))
+REPO = ROOT / "Ming-3.0-tts"
+MODEL = ROOT / "Ming-3.0-tts-0.5b"
+BUNDLE = ROOT / "webfm-eval"
+OUT = ROOT / "whisper_probe"
+SEED = 20261031
+
+LINES = [
+    "别急，先呼吸。",
+    "后来店关了，歌还在。",
+    "晚安，我们明天见。",
+]
+
+
+def log(msg: str) -> None:
+    print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)
+
+
+def patch_fp32() -> None:
+    for name in ("ming_tts.py", "modeling_bailingmm.py"):
+        p = REPO / name
+        s = p.read_text()
+        s = s.replace("torch_dtype=torch.bfloat16", "torch_dtype=torch.float32")
+        s = s.replace(".eval().to(torch.bfloat16).to(device)", ".eval().to(torch.float32).to(device)")
+        s = s.replace("dtype=torch.bfloat16", "dtype=torch.float32")
+        p.write_text(s)
+
+
+def voicing(y: np.ndarray, sr: int) -> float:
+    """Periodicity of the strongest frames. Low means breathy, no pitch."""
+    win, hop = int(0.04 * sr), int(0.02 * sr)
+    if len(y) < win * 2:
+        return 0.0
+    frames = np.lib.stride_tricks.sliding_window_view(y, win)[::hop]
+    energy = (frames ** 2).mean(1)
+    idx = np.argsort(energy)[-max(1, int(len(energy) * 0.3)):]
+    lo, hi = int(sr / 400), int(sr / 70)
+    peaks = []
+    for f in frames[idx]:
+        f = f - f.mean()
+        ac = np.correlate(f, f, "full")[win - 1:]
+        if ac[0] <= 0:
+            continue
+        peaks.append(float(ac[lo:hi].max() / ac[0]))
+    return float(np.median(peaks)) if peaks else 0.0
+
+
+def main() -> None:
+    import torch
+    import soundfile as sf
+    from transformers import pipeline
+
+    OUT.mkdir(parents=True, exist_ok=True)
+    sys.path.insert(0, str(REPO))
+    patch_fp32()
+    from ming_tts import MingTTS
+    sys.path.insert(0, str(BUNDLE / "eval"))
+    from score import load_t2s, score_row
+
+    table = load_t2s()
+    voices = json.loads((BUNDLE / "whisper/manifest.json").read_text())["voices"]
+    log(f"{len(voices)} whisper candidates x {len(LINES)} lines")
+
+    tts = MingTTS(str(MODEL), device="cuda", text_normalization=False)
+    rows = []
+    for v in voices:
+        wav_path = str(BUNDLE / v["wav"])
+        item = dict(v)
+        ces, vs, ws = [], [], []
+        for i, text in enumerate(LINES, 1):
+            try:
+                torch.manual_seed(SEED + i)
+                w = tts.synthesize(task="TTS", text=text, language_id="Chinese",
+                                   prompt_wav_path=wav_path, prompt_text=v["prompt_text"],
+                                   use_spk_emb=True, raw_text=True,
+                                   max_decode_steps=max(32, min(400, int(len(text) * 3.5))))
+                y = w.cpu().numpy().reshape(-1)
+                sf.write(str(OUT / f"{v['slot']}_{i}.wav"), y, 44100)
+                vs.append(voicing(y, 44100))
+                ws.append(round(y.size / 44100, 2))
+                item[f"sec_{i}"] = ws[-1]
+            except Exception:
+                log(f"{v['slot']} line {i} failed\n{traceback.format_exc()}")
+        item["voicing_out"] = float(np.mean(vs)) if vs else 0.0
+        item["seconds"] = ws
+        rows.append(item)
+        log(f"{v['slot']} src_v={v['voicing']:.2f} out_v={item['voicing_out']:.2f} "
+            f"sec={ws}")
+
+    asr = pipeline("automatic-speech-recognition", model="openai/whisper-small",
+                   device=0, torch_dtype=torch.float16)
+
+    for r in rows:
+        ces = []
+        for i, text in enumerate(LINES, 1):
+            p = OUT / f"{r['slot']}_{i}.wav"
+            if not p.exists():
+                continue
+            y, _ = sf.read(p, dtype="float32")
+            out = asr({"array": y.astype(np.float32), "sampling_rate": 44100},
+                      generate_kwargs={"language": "zh", "task": "transcribe"})
+            hyp = (out.get("text", "") if isinstance(out, dict) else str(out)).strip()
+            s = score_row({"text": text, "hyp_text": hyp}, table)
+            ces.append(s["cer_norm"])
+            r[f"hyp_{i}"] = hyp
+        r["cer"] = float(np.mean(ces)) if ces else 1.0
+        r["ok"] = bool(r["cer"] < 0.35 and r["voicing_out"] < 0.7)
+        log(f"{r['slot']} cer={r['cer']:.3f} v_out={r['voicing_out']:.2f} "
+            f"{'PASS' if r['ok'] else 'fail'}  {r.get('hyp_1','')[:24]}")
+
+    rows.sort(key=lambda r: (not r["ok"], r["cer"]))
+    (OUT / "metrics.json").write_text(json.dumps(rows, ensure_ascii=False, indent=1))
+    log("\n-- ranking --")
+    for r in rows:
+        log(f"  {'✓' if r['ok'] else '✗'} cer={r['cer']:.2f} v {r['voicing']:.2f}->"
+            f"{r['voicing_out']:.2f}  {r['artist'][:14]:<14} {r['prompt_text'][:18]}")
+
+
+if __name__ == "__main__":
+    main()