Blame for kaggle/voice_mining.py

"""Mine replacement prompt voices from DownFlow/ASMR.

The shipped voices failed as Ming prompts because their transcripts do not
describe their audio: v04's promptText covers ~2.8 s of speech while the clip
runs 5.98 s, so the clone learns a broken text/audio alignment that every
generated line inherits (teacher CER 0.83 versus 0.08 for a good prompt).

Two rules here:

1. every candidate is cropped on a phrase boundary and transcribed from the
   crop itself, never from the parent 30 s clip's transcript;
2. every candidate is ranked by the only measure that matters: bake it as a
   prompt, synthesise three fixed sentences with the teacher, transcribe,
   count character errors.

Stages:
  STAGE=harvest  CPU. stream, filter, crop, re-transcribe, score the signal
  STAGE=probe    GPU. bake each candidate, run the teacher probe, rank
  STAGE=all      both

Outputs: voices/{candidates/, candidates.json, probe.json, best.json}
"""
from __future__ import annotations

import hashlib
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"
WORK = ROOT / "voices"
CAND = WORK / "candidates"

DATASET = "DownFlow/ASMR"
N_ARTISTS = int(os.environ.get("N_ARTISTS", "20"))
N_PER_ARTIST = int(os.environ.get("N_PER_ARTIST", "6"))
CROP_S = float(os.environ.get("CROP_S", "6.0"))
MIN_SPEECH_S = float(os.environ.get("MIN_SPEECH_S", "4.5"))
STAGE = os.environ.get("STAGE", "all")
TOP_K = int(os.environ.get("TOP_K", "8"))
MAX_SCAN = int(os.environ.get("MAX_SCAN", "6000"))

PROBE_TEXTS = [
    "灯关小一点，我们慢慢来。",
    "雨从傍晚开始下，一直没停，窗户上蒙了一层薄薄的水汽。",
    "睡不着的夜里，时间会走得特别慢。",
]

BLOCK = [
    "操", "肏", "做爱", "阴茎", "阴道", "奶子", "龟头", "高潮", "自慰", "口交",
    "性爱", "色情", "淫", "骚", "舔我",
    "fuck", "cum", "penis", "pussy", "cock", "orgasm", "horny", "dick",
]

for _p in (WORK, CAND):
    _p.mkdir(parents=True, exist_ok=True)


def log(msg: str) -> None:
    print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)


def is_sfw(text: str) -> bool:
    t = (text or "").lower()
    return not any(w in t for w in BLOCK)


def frame_rms(y: np.ndarray, sr: int, hop_s: float = 0.01, win_s: float = 0.025) -> np.ndarray:
    hop, win = int(sr * hop_s), int(sr * win_s)
    if len(y) < win:
        return np.zeros(1, dtype=np.float32)
    n = 1 + (len(y) - win) // hop
    idx = np.arange(win)[None, :] + hop * np.arange(n)[:, None]
    return np.sqrt((y[idx] ** 2).mean(axis=1) + 1e-12)


def speech_mask(rms: np.ndarray) -> np.ndarray:
    floor = float(np.percentile(rms, 10))
    mask = rms > max(floor * 3.0, 1e-4)
    gap = 20                                  # 200 ms at a 10 ms hop
    i = 0
    while i < len(mask):
        if mask[i]:
            i += 1
            continue
        j = i
        while j < len(mask) and not mask[j]:
            j += 1
        if i > 0 and (j - i) < gap and j < len(mask):
            mask[i:j] = True
        i = j
    return mask


def crop_on_phrase(y: np.ndarray, sr: int):
    """Pick a CROP_S window beginning at a phrase onset. Returns (audio, stats)."""
    rms = frame_rms(y, sr)
    mask = speech_mask(rms)
    runs, start = [], (0 if mask[0] else None)
    for e in np.flatnonzero(np.diff(mask.astype(np.int8))):
        if start is None:
            start = e + 1
        else:
            runs.append((start, int(e) + 1))
            start = None
    if start is not None:
        runs.append((start, len(mask)))

    per_10ms = int(sr * 0.01)
    need = int(MIN_SPEECH_S * 100)
    runs = [r for r in runs if r[1] - r[0] >= need]
    if not runs:
        return None, {"reject": "no_phrase"}

    s, e = max(runs, key=lambda r: float(rms[r[0]:r[1]].mean()))
    n = int(CROP_S * 100)
    if e - s < n:
        s = max(s, e - n)
    else:
        e = s + n

    seg = np.asarray(y[s * per_10ms: e * per_10ms], dtype=np.float32)
    if seg.size < per_10ms * 100 * 0.8:
        return None, {"reject": "short_crop"}
    peak = float(np.abs(seg).max())
    if peak > 0:
        seg = seg * (0.90 / peak)

    noise = rms[~mask] if (~mask).any() else rms
    speech = rms[s:e][mask[s:e]]
    snr = 10 * np.log10(max(1e-8, float(speech.mean()) ** 2 / max(1e-8, float(noise.mean()) ** 2)))
    stats = {
        "speech_ratio": float(mask[s:e].mean()),
        "snr_db": float(snr),
        "clip_ratio": float((np.abs(seg) > 0.995).mean()),
    }
    return seg, stats


def spectral_flatness(y: np.ndarray, sr: int) -> float:
    win = int(0.025 * sr)
    if len(y) < win:
        return 1.0
    frames = np.lib.stride_tricks.sliding_window_view(y, win)[:: int(0.01 * sr)]
    spec = np.abs(np.fft.rfft(frames * np.hanning(win), axis=1)) + 1e-10
    return float(np.median(np.exp(np.log(spec).mean(axis=1)) / spec.mean(axis=1)))


def speech_modulation(y: np.ndarray, sr: int) -> float:
    """Share of envelope energy in the 2-8 Hz syllabic band.

    Speech has a syllabic rhythm; rustling, tapping and room tone do not. This
    is the physical counterpart to the transcript: whisper-small happily
    invents a caption over pure noise ("字幕by索兰娅"), so the transcript alone
    cannot tell us whether anyone is talking.
    """
    hop = int(0.01 * sr)
    if len(y) < hop * 50:
        return 0.0
    n = len(y) // hop
    env = np.sqrt((y[: n * hop].reshape(n, hop) ** 2).mean(axis=1) + 1e-12)
    env = env - env.mean()
    spec = np.abs(np.fft.rfft(env))
    freqs = np.fft.rfftfreq(len(env), 0.01)
    total = spec.sum() + 1e-12
    return float(spec[(freqs >= 2.0) & (freqs <= 8.0)].sum() / total)


ASR_ARTIFACTS = ("字幕", "by索", "谢谢观看", "请订阅", "amara", "subtitles", "感谢观看")


def repeated_run(text: str) -> float:
    """Share of the transcript taken by its longest repeated character run."""
    if not text:
        return 1.0
    best, cur = 1, 1
    for a, b in zip(text, text[1:]):
        cur = cur + 1 if a == b else 1
        best = max(best, cur)
    return best / len(text)


def build_asr():
    import torch
    from transformers import pipeline

    pipe = pipeline("automatic-speech-recognition", model="openai/whisper-small",
                    device=0, torch_dtype=torch.float16)

    def asr(y: np.ndarray, sr: int) -> str:
        out = pipe({"array": y.astype(np.float32), "sampling_rate": sr},
                   generate_kwargs={"language": "zh", "task": "transcribe"})
        return (out.get("text", "") if isinstance(out, dict) else str(out)).strip()

    return asr


def harvest(asr) -> list[dict]:
    import soundfile as sf
    from datasets import load_dataset

    ds = load_dataset(DATASET, split="train", streaming=True)
    per_artist: dict[str, list[dict]] = {}
    scanned = 0
    t0 = time.time()

    for row in ds:
        scanned += 1
        if scanned > MAX_SCAN or len(per_artist) >= N_ARTISTS:
            break
        creator = row.get("creator") or "unknown"
        if len(per_artist.get(creator, [])) >= N_PER_ARTIST:
            continue
        if creator not in per_artist and len(per_artist) >= N_ARTISTS:
            continue
        if row.get("nsfw") or not row.get("has_verbal_speech") or row.get("hallucinated"):
            continue
        if not is_sfw(row.get("transcript", "")):
            continue

        try:
            audio = row["audio"]
            y = np.asarray(audio["array"], dtype=np.float32).reshape(-1)
            sr = int(audio["sampling_rate"])
        except Exception:
            continue

        seg, st = crop_on_phrase(y, sr)
        if seg is None:
            continue
        flat = spectral_flatness(seg, sr)
        mod = speech_modulation(seg, sr)
        st["flatness"] = flat
        st["modulation"] = mod
        if st["snr_db"] < 12 or st["clip_ratio"] > 1e-3 or st["speech_ratio"] < 0.55 or flat > 0.35:
            continue
        # Do not gate on `modulation`. Measured on clips a listener confirmed
        # contain whisper speech versus clips that are pure ASMR noise, the two
        # overlap almost exactly (0.239 vs 0.230), because whispering has weak
        # syllabic rhythm. It is recorded for diagnostics only. The reliable
        # discriminator is the round-trip probe CER: noise-only prompts score
        # above 1.0, a usable whisper prompt scores 0.22.

        text = asr(seg, sr)
        if not (6 <= len(text) <= 70) or repeated_run(text) > 0.3 or not is_sfw(text):
            continue
        if any(a in text for a in ASR_ARTIFACTS):
            continue
        # A hallucinated caption over noise tends to be far shorter or longer
        # than anyone can actually say in six seconds.
        rate = len(text) / (len(seg) / sr)
        if not (2.5 <= rate <= 9.0):
            continue

        idx = len(per_artist.get(creator, []))
        tag = hashlib.md5(creator.encode("utf-8")).hexdigest()[:4]
        cid = f"{tag}_{idx}"
        path = CAND / f"{cid}.wav"
        sf.write(str(path), seg, sr)
        per_artist.setdefault(creator, []).append({
            "cid": cid, "artist": creator, "wav": str(path),
            "prompt_text": text, "duration_s": round(len(seg) / sr, 2), **st,
            "source_file": row.get("source_file"), "clip_index": row.get("clip_index"),
        })
        if len(per_artist[creator]) == 1:
            log(f"artist {len(per_artist)}: {creator[:40]} first candidate ok")

    out = [c for cl in per_artist.values() for c in cl]
    (WORK / "candidates.json").write_text(json.dumps(out, ensure_ascii=False, indent=2))
    log(f"harvest: {len(per_artist)} artists, {len(out)} candidates, scanned {scanned} "
        f"rows in {time.time()-t0:.0f}s")
    return out


def hf_ratio(y: np.ndarray, sr: int, cut: int = 8000) -> float:
    """Energy share above `cut`. A 16 kHz source cannot produce it, so a dull
    clone shows up here even when CER looks fine."""
    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 patch_fp32() -> None:
    """T4 has no bf16 and the repo defaults to it."""
    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 probe(cands: list[dict], asr) -> list[dict]:
    import torch

    sys.path.insert(0, str(REPO))
    patch_fp32()
    from ming_tts import MingTTS
    import tts_tasks

    sys.path.insert(0, str(ROOT / "webfm-eval/eval"))
    from score import load_t2s, score_row

    table = load_t2s()
    tts = MingTTS(str(MODEL), device="cuda", text_normalization=False)
    mm = tts.model
    device = next(mm.parameters()).device
    dtype = next(mm.parameters()).dtype
    _ = tts_tasks.build_tts_prompt("Chinese")

    results = []
    for i, c in enumerate(cands):
        try:
            t0 = time.time()
            torch.manual_seed(1234)
            lat, spk = tts.prepare_prompt(c["wav"], use_spk_emb=True)
            lat = lat.to(device=device, dtype=dtype)
            ces, hfs = [], []
            for text in PROBE_TEXTS:
                wav = tts.synthesize(
                    task="TTS", text=text, language_id="Chinese",
                    prompt_wav_path=c["wav"], prompt_text=c["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)
                ces.append(score_row({"text": text, "hyp_text": asr(arr, 44100)}, table)["cer_norm"])
                hfs.append(hf_ratio(arr, 44100))
            row = dict(c, cer=float(np.mean(ces)), cer_each=ces,
                       hf=float(np.mean(hfs)), wall=round(time.time() - t0, 1))
            # A prompt whose round trip scores above 0.5 is not a voice, it is
            # either a bad transcript or a clip with no speech in it. Both show
            # up here as garbled output, and both are unusable downstream.
            row["keep"] = bool(row["cer"] < 0.5)
            results.append(row)
            log(f"[{i+1}/{len(cands)}] {c['cid']} cer={row['cer']:.3f} hf={row['hf']:.4f} "
                f"snr={c['snr_db']:.1f} {c['prompt_text'][:16]}")
        except Exception:
            log(f"{c['cid']} probe failed\n{traceback.format_exc()}")

    results.sort(key=lambda r: (r["cer"], -r["hf"]))
    (WORK / "probe.json").write_text(json.dumps(results, ensure_ascii=False, indent=2))
    return results


def select(results: list[dict]) -> list[dict]:
    best: dict[str, dict] = {}
    for r in results:
        best.setdefault(r["artist"], r)
    ranked = [r for r in sorted(best.values(), key=lambda r: r["cer"])
              if r.get("keep", True)][:TOP_K]
    (WORK / "best.json").write_text(json.dumps(ranked, ensure_ascii=False, indent=2))
    log("\n-- best per artist --")
    for r in ranked:
        log(f"  {r['cer']:.3f}  hf={r['hf']:.4f}  {r['artist'][:34]:<34} "
            f"{r['duration_s']}s  {r['prompt_text'][:20]}")
    return ranked


def main() -> None:
    asr = build_asr()
    if STAGE in ("harvest", "all"):
        cands = harvest(asr)
    else:
        cands = json.loads((WORK / "candidates.json").read_text())
    if STAGE in ("probe", "all"):
        select(probe(cands, asr))


if __name__ == "__main__":
    main()