app.py

import gradio as gr
import torch
import os
import tempfile
import time
from TTS.api import TTS

# Global variable to store the TTS model
model = None

def load_model():
    global model
    try:
        # Load model on first run
        if model is None:
            print("Loading XTTS model... (this may take a few minutes)")
            model = TTS("tts_models/multilingual/multi-dataset/xtts_v2")
            print("Model loaded successfully!")
        return True
    except Exception as e:
        print(f"Error loading model: {str(e)}")
        return False

def clone_voice(text, audio_file, progress=gr.Progress()):
    """
    Generate speech using XTTS with the uploaded reference audio
    """
    if not text or not audio_file:
        return None, "Please provide both text and a voice sample."
    
    progress(0.1, desc="Loading model...")
    
    # Try to load the model (if not already loaded)
    model_loaded = load_model()
    if not model_loaded:
        return None, "Failed to load the TTS model. This might be due to memory constraints."
    
    progress(0.3, desc="Processing voice sample...")
    
    try:
        # Generate speech with the cloned voice
        progress(0.6, desc="Generating speech (this may take a while on CPU)...")
        
        # Create a temporary directory for output
        temp_dir = tempfile.gettempdir()
        output_path = os.path.join(temp_dir, "xtts_output.wav")
        
        # Use the model to generate speech
        start_time = time.time()
        model.tts_to_file(
            text=text,
            file_path=output_path,
            speaker_wav=audio_file,
            language="en"
        )
        generation_time = time.time() - start_time
        
        progress(1.0, desc="Done!")
        return output_path, f"Generation completed in {generation_time:.1f} seconds."
        
    except Exception as e:
        error_message = str(e)
        if "CUDA" in error_message:
            return None, "CUDA error: This Space is running on CPU. Please keep your text short for better performance."
        elif "memory" in error_message.lower():
            return None, "Memory error: The model ran out of memory. Try with shorter text or a smaller voice sample."
        else:
            return None, f"Error generating speech: {error_message}"

# Create the Gradio interface
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# XTTS Voice Cloning")
    gr.Markdown("""
    This app uses XTTS v2 to clone a voice from a sample and generate speech with that voice.
    
    ⚠️ **CPU Warning**: This is running on a CPU, so generation may take 1-3 minutes. 
    Keep your text short (1-2 sentences) for best results.
    """)
    
    with gr.Row():
        with gr.Column():
            text_input = gr.Textbox(
                label="Text to speech",
                placeholder="Enter text to be spoken in the cloned voice (keep it short)...",
                lines=3
            )
            
            audio_input = gr.Audio(
                label="Voice reference (5-10 second sample)", 
                type="filepath"
            )
            
            with gr.Row():
                clear_btn = gr.Button("Clear")
                submit_btn = gr.Button("Generate", variant="primary")
        
        with gr.Column():
            output_audio = gr.Audio(label="Generated Speech")
            output_message = gr.Textbox(label="Status")
    
    # Set up interactions
    submit_btn.click(
        fn=clone_voice,
        inputs=[text_input, audio_input],
        outputs=[output_audio, output_message]
    )
    
    clear_btn.click(
        fn=lambda: (None, None, ""),
        inputs=[],
        outputs=[text_input, output_audio, output_message]
    )
    
    gr.Markdown("""
    ## Tips for better results:
    1. Use a clear voice recording with minimal background noise
    2. Keep the sample 5-10 seconds long
    3. For faster generation, keep your text very short (1-2 sentences max)
    4. Be patient! CPU generation takes time
    
    ## License
    This app uses the XTTS v2 model under the Coqui Public Model License for non-commercial use only.
    """)

    # Comment out the problematic line - we don't need it
    # demo.load(lambda: None, None, None, _js="() => {console.log('Loading model in the background...')}")

# Launch the app
demo.queue().launch()