tts.py

import os
import torch
import torchaudio
import spaces  # Import spaces module for Zero-GPU
from tortoise.api import TextToSpeech
from tortoise.utils.audio import load_audio

# Create cache/output directory
os.makedirs("outputs", exist_ok=True)

# Create a global TTS model instance
tts_model = None

@spaces.GPU  # Add spaces.GPU decorator for Zero-GPU support
async def generate_speech(text, voice_preset="random", voice_file_path=None):
    global tts_model
    
    try:
        # Initialize the model if not already initialized
        if tts_model is None:
            print("Initializing Tortoise-TTS model...")
            tts_model = TextToSpeech(use_deepspeed=torch.cuda.is_available())
            print(f"Model initialized. Using device: {next(tts_model.autoregressive.parameters()).device}")
        
        # Process voice sample if provided
        voice_samples = None
        if voice_file_path and os.path.exists(voice_file_path):
            print(f"Loading voice from {voice_file_path}")
            voice_samples, _ = load_audio(voice_file_path, 22050)
            voice_samples = [voice_samples]
            voice_preset = None
        
        # Generate speech
        print(f"Generating speech for text: {text[:50]}...")
        output_filename = f"outputs/tts_output_{hash(text) % 10000}.wav"
        
        gen = tts_model.tts_with_preset(
            text,
            voice_samples=voice_samples,
            preset=voice_preset
        )
        
        # Save the generated audio
        torchaudio.save(output_filename, gen.squeeze(0).cpu(), 24000)
        print(f"Speech generated and saved to {output_filename}")
        
        # Return the filename and audio data
        return output_filename, (24000, gen.squeeze(0).cpu())
        
    except Exception as e:
        print(f"Error generating speech: {str(e)}")
        raise