backend/voice/stt.py

# backend/voice/stt.py
# Real STT using faster-whisper (local) with Deepgram fallback
import os

APP_DATA_DIR = os.environ.get("JARVIS_APP_DATA_DIR", ".")

class STTPipeline:
    def __init__(self, model_size="large-v3"):
        import torch
        from faster_whisper import WhisperModel
        # Downloads real Whisper model to {app_data_dir}/models/whisper/
        self.model = WhisperModel(model_size,
                                   device="cuda" if torch.cuda.is_available() else "cpu",
                                   compute_type="float16" if torch.cuda.is_available() else "int8",
                                   download_root=f"{APP_DATA_DIR}/models/whisper")

    async def transcribe(self, audio_bytes: bytes) -> tuple[str, str]:
        # language=None forces auto-detection
        segments, info = self.model.transcribe(audio_bytes, beam_size=5, language=None)
        
        text = " ".join(s.text for s in segments)
        
        detected_lang = info.language
        if info.duration < 2.0 and hasattr(self, "last_known_language"):
            detected_lang = self.last_known_language
        else:
            self.last_known_language = detected_lang
            
        import gc, torch
        del segments
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
            
        return text, detected_lang

    def unload(self):
        import gc
        import torch
        if hasattr(self, 'model'):
            del self.model
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()