File size: 4,682 Bytes
bf64381 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 | """Clone the eight mined voices with Ming and render a sample for each.
The point is comparison: the listening page already has the raw 6 s crop, this
adds what the model actually produces from it, on the same two sentences, so
timbre can be judged before any voice joins the roster.
Reads webfm-eval/mined/manifest.json, writes voices_clone/{slot}_1.wav,
voices_clone/{slot}_2.wav and voices_clone/metrics.json.
Note the prompt transcripts come from adult-roleplay ASMR clips. Only the timbre
is cloned; the prompt text is conditioning, never spoken in the product. Source
provenance still needs to hold up in the licence audit before any of these ship.
"""
import json
import os
import sys
import time
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 / "voices_clone"
TEXTS = [
"灯关小一点，我们慢慢来。",
"睡不着的夜里，时间会走得特别慢。",
]
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"):
path = REPO / name
s = path.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")
path.write_text(s)
def hf_ratio(y: np.ndarray, sr: int, cut: int = 8000) -> float:
spec = np.abs(np.fft.rfft(y)) ** 2
freqs = np.fft.rfftfreq(len(y), 1 / sr)
return float(spec[freqs > cut].sum() / (spec.sum() + 1e-12))
def main() -> None:
import torch
import torchaudio
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()
from transformers import pipeline
asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-small",
device=0, torch_dtype=torch.float16)
def asr(y: np.ndarray) -> str:
out = asr_pipe({"array": y.astype(np.float32), "sampling_rate": 44100},
generate_kwargs={"language": "zh", "task": "transcribe"})
return (out.get("text", "") if isinstance(out, dict) else str(out)).strip()
tts = MingTTS(str(MODEL), device="cuda", text_normalization=False)
voices = json.loads((BUNDLE / "mined/manifest.json").read_text())["voices"]
log(f"{len(voices)} voices, {len(TEXTS)} sentences each")
rows = []
for v in voices:
wav_path = BUNDLE / v["wav"]
entry = dict(v)
for i, text in enumerate(TEXTS, 1):
try:
t0 = time.time()
torch.manual_seed(1234)
wav = tts.synthesize(
task="TTS", text=text, language_id="Chinese",
prompt_wav_path=str(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))),
)
arr = wav.cpu().numpy().reshape(-1)
path = OUT / f"{v['slot']}_{i}.wav"
torchaudio.save(str(path), torch.from_numpy(arr).unsqueeze(0), 44100)
s = score_row({"text": text, "hyp_text": asr(arr)}, table)
entry[f"cer_{i}"] = s["cer_norm"]
entry[f"asr_{i}"] = s["hyp_text"]
entry[f"hf_{i}"] = hf_ratio(arr, 44100)
entry[f"sec_{i}"] = round(arr.size / 44100, 2)
log(f"{v['slot']} #{i} cer={s['cer_norm']:.3f} hf={entry[f'hf_{i}']:.4f} "
f"{arr.size/44100:.1f}s {time.time()-t0:.0f}s")
except Exception as exc:
entry[f"error_{i}"] = f"{type(exc).__name__}: {exc}"
log(f"{v['slot']} #{i} failed: {exc}")
entry["cer"] = float(np.mean([entry[k] for k in entry if k.startswith("cer_")] or [9]))
entry["hf"] = float(np.mean([entry[k] for k in entry if k.startswith("hf_")] or [0]))
rows.append(entry)
rows.sort(key=lambda r: r["cer"])
(OUT / "metrics.json").write_text(json.dumps(rows, ensure_ascii=False, indent=1))
log("\n-- by CER --")
for r in rows:
log(f" {r['cer']:.3f} hf={r['hf']:.4f} {r['slot']} {r['artist'][:16]:<16} "
f"mine_cer={r['mine_cer']:.3f}")
if __name__ == "__main__":
main()
|