assistant/stt.py

from __future__ import annotations

from functools import lru_cache

from faster_whisper import WhisperModel

from config import CONFIG


@lru_cache(maxsize=1)
def _model():
    """Load the whisper model once and cache it."""
    return WhisperModel(
        CONFIG.whisper_size,
        device="cpu",
        compute_type=CONFIG.whisper_compute,
    )


def transcribe(audio_path: str) -> str:
    """Transcribe a recording to text in the configured language."""
    if not audio_path:
        return ""
    segments, _info = _model().transcribe(
        audio_path,
        language=CONFIG.language,
        vad_filter=True,
    )
    return " ".join(seg.text.strip() for seg in segments).strip()