watch_sceen_ai/wsai/backends/whisper_worker.py

"""Persistent faster-whisper STT worker.

faster-whisper (ctranslate2) lives in its own Python (whisper312); loading the
model takes seconds, so we load it ONCE here and then serve transcription
requests over stdin/stdout. This process is launched with the whisper312
interpreter by wsai.backends.whisper.WhisperSTT.

Like the MeloTTS worker, model/backend chatter could corrupt the JSON protocol,
so on startup we split the streams: a private duplicate of the original stdout
carries the protocol, and fd 1 is redirected to fd 2 so any library print lands
on stderr instead (where the parent drains it for diagnostics).

Protocol (one JSON object per line, on the protocol channel):
    <- {"wav": "/abs/path.wav", "language": "ko"}
    -> {"ok": true, "text": "...", "language": "ko", "ms": 123}
    -> {"ok": false, "error": "..."}
On startup, once the model is ready, it emits exactly one line:
    -> {"ready": true, "ms": <load-ms>, "device": "cpu", "model": "small"}
"""

import json
import os
import sys
import time

# Whisper's classic Korean hallucinations on silence/noise/keyboard clatter —
# it "hears" video-outro boilerplate. Drop these when the whole utterance is one
# of them AND the segment looked like non-speech, so a real "감사합니다" survives.
_HALLUCINATIONS = {
    "감사합니다", "고맙습니다", "감사합니다.", "고맙습니다.",
    "시청해주셔서 감사합니다", "시청해 주셔서 감사합니다", "끝까지 시청해주셔서 감사합니다",
    "구독과 좋아요 부탁드립니다", "구독 좋아요 부탁드립니다", "다음 영상에서 만나요",
    "다음 시간에 만나요", "안녕히 계세요",
}


def _norm(t: str) -> str:
    return t.strip().rstrip(" .!?…~").strip()


def _guard_hallucination(text: str, worst_no_speech: float) -> str:
    """Blank out a lone known-hallucination phrase when the audio was probably
    not speech (high no_speech_prob)."""
    n = _norm(text)
    if not n:
        return ""
    if n in {_norm(h) for h in _HALLUCINATIONS} and worst_no_speech > 0.5:
        return ""
    return text

# Split protocol from library noise BEFORE importing anything heavy.
_proto = os.fdopen(os.dup(1), "w", buffering=1)  # private copy of real stdout
os.dup2(2, 1)  # fd1 -> stderr, so stray library prints don't hit the protocol


def _emit(obj: dict) -> None:
    _proto.write(json.dumps(obj, ensure_ascii=False) + "\n")
    _proto.flush()


def _log(*a):
    print(*a, file=sys.stderr, flush=True)


def main() -> None:
    model_name = os.environ.get("WSAI_WHISPER_MODEL", "small")
    requested = os.environ.get("WSAI_WHISPER_DEVICE", "auto")  # cpu | cuda | auto
    default_lang = os.environ.get("WSAI_WHISPER_LANGUAGE", "ko") or None

    from faster_whisper import WhisperModel  # heavy import; only in whisper venv

    def _has_cuda() -> bool:
        try:
            import ctranslate2

            return ctranslate2.get_cuda_device_count() > 0
        except Exception:
            return False

    device = requested
    if requested == "auto":
        device = "cuda" if _has_cuda() else "cpu"

    def _compute_for(dev: str) -> str:
        # int8 on CPU keeps a small model fast; float16 is the usual CUDA choice.
        return os.environ.get(
            "WSAI_WHISPER_COMPUTE", "int8" if dev == "cpu" else "float16"
        )

    t0 = time.monotonic()
    try:
        model = WhisperModel(model_name, device=device, compute_type=_compute_for(device))
    except Exception as exc:
        # CUDA picked but unusable (missing libs, OOM): fall back to CPU rather
        # than leaving the whole voice loop dead.
        if device == "cuda":
            _log(f"[whisper_worker] CUDA load failed ({exc}); falling back to CPU")
            device = "cpu"
            model = WhisperModel(model_name, device=device, compute_type=_compute_for(device))
        else:
            raise
    compute = _compute_for(device)
    load_ms = int((time.monotonic() - t0) * 1000)

    # Warm up before signalling ready: the first CUDA transcribe pays a large
    # lazy cost (kernel autotune), which would slow the first real utterance.
    # Run a dummy transcribe on 1s of silence here so "ready" means "hot".
    warmup_ms = None
    try:
        import numpy as np

        w = time.monotonic()
        segs, _ = model.transcribe(np.zeros(16000, dtype=np.float32), language=default_lang)
        for _ in segs:  # segments are lazy; drain to force the actual compute
            pass
        warmup_ms = int((time.monotonic() - w) * 1000)
    except Exception as exc:
        _log(f"[whisper_worker] warmup skipped: {exc}")

    _emit({"ready": True, "ms": load_ms, "device": device, "model": model_name, "warmup_ms": warmup_ms})
    _log(f"[whisper_worker] {model_name} ready in {load_ms} ms on {device}/{compute} (warmup {warmup_ms} ms)")

    for line in sys.stdin:
        line = line.strip()
        if not line:
            continue
        try:
            req = json.loads(line)
            wav = req["wav"]
            language = req.get("language", default_lang)
            s = time.monotonic()
            segments, info = model.transcribe(
                wav,
                language=language,
                beam_size=int(req.get("beam_size", 5)),
                # VAD strips non-speech (keyboard clatter, room noise, silence)
                # before decoding, which both improves accuracy and kills most
                # hallucinations. speech_pad_ms keeps a little lead/trail so soft
                # first/last words aren't clipped.
                vad_filter=bool(req.get("vad_filter", True)),
                vad_parameters=dict(min_silence_duration_ms=300, speech_pad_ms=250),
                # Don't feed the previous text back in — that's what makes Whisper
                # loop/hallucinate. Temperature fallback + thresholds reject
                # low-confidence (noisy/quiet) decodes instead of inventing words.
                condition_on_previous_text=False,
                temperature=[0.0, 0.2, 0.4, 0.6],
                no_speech_threshold=0.6,
                log_prob_threshold=-1.0,
                compression_ratio_threshold=2.4,
            )
            parts, worst_ns = [], 0.0
            for seg in segments:
                nsp = float(getattr(seg, "no_speech_prob", 0.0) or 0.0)
                alp = float(getattr(seg, "avg_logprob", 0.0) or 0.0)
                # Drop a segment that is almost certainly non-speech noise.
                if nsp > 0.8 and alp < -0.4:
                    continue
                worst_ns = max(worst_ns, nsp)
                parts.append(seg.text)
            text = _guard_hallucination("".join(parts).strip(), worst_ns)
            ms = int((time.monotonic() - s) * 1000)
            _emit({"ok": True, "text": text, "language": info.language, "ms": ms})
        except Exception as exc:  # keep the worker alive across bad requests
            _emit({"ok": False, "error": f"{type(exc).__name__}: {exc}"})
            _log(f"[whisper_worker] error: {exc}")


if __name__ == "__main__":
    main()