transcriber.py

"""Speech-to-text.

Primary path is Gemma 4's audio modality (one model for the whole pipeline). Gemma
caps audio at ~30s, so longer notes fall back to faster-whisper, which handles any
length. If the duration can't be probed, we try Gemma and fall back on any error.
"""

from __future__ import annotations

import threading

import gemma
import llama_backend
from config import (
    GEMMA_AUDIO_MAX_SEC,
    WHISPER_BEAM,
    WHISPER_COMPUTE,
    WHISPER_DEVICE,
    WHISPER_LANG,
    WHISPER_MODEL,
)

_whisper = None
_lock = threading.Lock()


def _duration(audio_path: str) -> float | None:
    """Best-effort audio length in seconds; None if it can't be determined."""
    try:
        import librosa

        return float(librosa.get_duration(path=audio_path))
    except Exception:
        return None


def _get_whisper():
    global _whisper
    if _whisper is None:
        with _lock:
            if _whisper is None:
                from faster_whisper import WhisperModel

                _whisper = WhisperModel(
                    WHISPER_MODEL, device=WHISPER_DEVICE, compute_type=WHISPER_COMPUTE,
                )
    return _whisper


def _whisper_transcribe(audio_path: str) -> str:
    model = _get_whisper()
    segments, _info = model.transcribe(
        audio_path,
        language=WHISPER_LANG,
        beam_size=WHISPER_BEAM,
        vad_filter=True,
    )
    return " ".join(seg.text.strip() for seg in segments).strip()


def transcribe(audio_path: str) -> str:
    """Transcribe an audio file to plain text. Returns '' for no/empty input.

    Order: llama.cpp (Gemma mtmd) -> Gemma transformers -> faster-whisper. The Gemma
    paths only run for clips within the 30s audio cap; longer notes go straight to
    whisper, which handles any length."""
    if not audio_path:
        return ""

    short = (dur := _duration(audio_path)) is None or dur <= GEMMA_AUDIO_MAX_SEC
    if short:
        # Primary: llama.cpp mtmd (Gemma audio encoder)
        try:
            text = llama_backend.transcribe_audio(audio_path)
            if text:
                return text
        except Exception:
            pass
        # Fallback: Gemma via transformers
        try:
            text = gemma.transcribe_audio(audio_path)
            if text:
                return text
        except Exception:
            pass
    # Last resort (and all long audio): faster-whisper
    return _whisper_transcribe(audio_path)


if __name__ == "__main__":  # quick CLI: python transcriber.py path/to/audio.m4a
    import sys

    if len(sys.argv) < 2:
        print("usage: python transcriber.py <audio-file>")
        raise SystemExit(1)
    print(transcribe(sys.argv[1]))