Voice MCP

"""Whisper transcription wrapper using faster-whisper.""" import sys import numpy as np from faster_whisper import WhisperModel # Global model cache _model: WhisperModel | None = None _model_name: str = "" def load_model(model_name: str = "small") -> WhisperModel: """Load Whisper model (cached after first load).""" global _model, _model_name if _model is not None and _model_name == model_name: return _model print(f"Loading Whisper model '{model_name}'...", file=sys.stderr) # Use CPU - GPU requires cuDNN which may not be installed _model = WhisperModel(model_name, device="cpu", compute_type="int8") _model_name = model_name print("Model loaded.", file=sys.stderr) return _model def transcribe(audio: np.ndarray, model_name: str = "small") -> dict: """ Transcribe audio using Whisper. Args: audio: numpy array of audio at 16kHz mono float32 model_name: Whisper model size (tiny, base, small, medium, large-v3) Returns: dict with 'text' and 'language' keys """ if len(audio) == 0: return {"text": "", "language": ""} model = load_model(model_name) # Whisper expects float32 audio normalized to [-1, 1] audio = audio.astype(np.float32) # Transcribe - faster-whisper handles audio of any length segments, info = model.transcribe(audio, beam_size=5) # Collect all segment texts text_parts = [] for segment in segments: text_parts.append(segment.text) full_text = "".join(text_parts).strip() return { "text": full_text, "language": info.language, }