services/stt_service.py

import torch
import torchaudio
from transformers import pipeline
import logging
import tempfile
import os

logger = logging.getLogger(__name__)

# Global STT pipeline
stt_pipeline = None

def load_stt_model():
    """Load the free Whisper model for speech-to-text"""
    global stt_pipeline
    try:
        logger.info("Loading Whisper STT model...")
        stt_pipeline = pipeline(
            "automatic-speech-recognition",
            model="openai/whisper-small",  # Free model
            device="cpu"  # Use CPU to avoid GPU requirements
        )
        logger.info("✓ Whisper STT model loaded successfully")
    except Exception as e:
        logger.error(f"✗ Failed to load Whisper model: {str(e)}")
        stt_pipeline = None

async def speech_to_text(audio_bytes: bytes, filename: str) -> str:
    """
    Convert audio bytes to text using free Whisper model.
    
    Args:
        audio_bytes: Raw audio file bytes
        filename: Name of the audio file
    
    Returns:
        Transcribed text
    """
    global stt_pipeline
    
    try:
        if stt_pipeline is None:
            load_stt_model()
            if stt_pipeline is None:
                raise Exception("STT model failed to load")
        
        logger.info(f"Converting audio to text using Whisper")
        
        # Save audio bytes to temporary file
        with tempfile.NamedTemporaryFile(delete=False, suffix='.wav') as temp_audio:
            temp_audio.write(audio_bytes)
            temp_audio_path = temp_audio.name
        
        try:
            # Transcribe using Whisper
            result = stt_pipeline(temp_audio_path)
            transcribed_text = result.get("text", "").strip()
            
            if not transcribed_text:
                transcribed_text = "Sorry, I couldn't understand the audio."
                
            logger.info(f"✓ STT successful: '{transcribed_text}'")
            return transcribed_text
            
        finally:
            # Clean up temporary file
            if os.path.exists(temp_audio_path):
                os.unlink(temp_audio_path)
                
    except Exception as e:
        logger.error(f"✗ STT failed: {str(e)}")
        raise Exception(f"Speech-to-text conversion failed: {str(e)}")