app.py

import gradio as gr
# Import alias module before outetts to setup whisper redirection
import alias as _alias
import outetts
import json
import tempfile
import hashlib
import os
from typing import Optional
from outetts.models.info import MODEL_INFO
from outetts.utils import helpers
from huggingface_hub import hf_hub_download
import torch
from transformers import BitsAndBytesConfig
import spaces

# Available OuteTTS models based on the documentation
MODELS = {v.value: v for _, v in outetts.Models.__members__.items()}

MODEL_QUANTIZATION = {
    outetts.Models.VERSION_0_1_SIZE_350M: outetts.LlamaCppQuantization.Q8_0,
    outetts.Models.VERSION_0_2_SIZE_500M: outetts.LlamaCppQuantization.Q8_0,
    outetts.Models.VERSION_0_3_SIZE_500M: outetts.LlamaCppQuantization.Q8_0,
}

# Cache for speaker profiles to avoid re-transcribing the same audio
speaker_cache = {}

def get_file_hash(file_path):
    """Calculate MD5 hash of a file for caching purposes."""
    hash_md5 = hashlib.md5()
    with open(file_path, "rb") as f:
        for chunk in iter(lambda: f.read(4096), b""):
            hash_md5.update(chunk)
    return hash_md5.hexdigest()

def try_ggml_model(model: outetts.Models, backend: outetts.Backend, quantization: outetts.LlamaCppQuantization):
    model_config = MODEL_INFO[model]
    repo = f"OuteAI/{model.value}-GGUF"
    filename = f"{model.value}-{quantization.value}.gguf"
    model_path = hf_hub_download(
        repo_id=repo,
        filename=filename,
        local_dir=os.path.join(helpers.get_cache_dir(), "gguf"),
        local_files_only=False
    )
    return outetts.ModelConfig(
        model_path=model_path,
        tokenizer_path=f"OuteAI/{model.value}",
        backend=backend,
        n_gpu_layers=99,
        verbose=False,
        device=None,
        dtype=None,
        additional_model_config={},
        audio_codec_path=None,
        **model_config
    )

@spaces.GPU
def get_interface(model_name: str):
    """Get interface instance for the model (no caching to avoid CUDA memory issues)."""
    model = MODELS[model_name]

    has_cuda = torch.cuda.is_available()
    if has_cuda:
        model_config = MODEL_INFO[model]
        config = outetts.ModelConfig(
            model_path=f"OuteAI/{model_name}",
            tokenizer_path=f"OuteAI/{model_name}",
            backend=outetts.Backend.HF,
            additional_model_config={
                "device_map": "auto" if has_cuda else "cpu",
                "quantization_config": BitsAndBytesConfig(
                    load_in_8bit=True
                ) if has_cuda else None,
            },
            **model_config
        )
    else:
        quantization = MODEL_QUANTIZATION.get(model, outetts.LlamaCppQuantization.Q6_K)
        config = try_ggml_model(model, outetts.Backend.LLAMACPP, quantization)
    
    # Initialize the interface
    interface = outetts.Interface(config=config)
    return interface

def get_or_create_speaker(interface, audio_file):
    """Get speaker from cache or create new one if not cached."""
    # Calculate file hash for caching
    file_hash = get_file_hash(audio_file)
    cache_key = f"{interface.config.interface_version}_{file_hash}"
    
    # Check if speaker profile is already cached
    if cache_key in speaker_cache:
        print(f"✅ Using cached speaker profile for {os.path.basename(audio_file)}")
        return speaker_cache[cache_key]
    
    # Create new speaker profile
    print(f"🔄 Creating new speaker profile for {os.path.basename(audio_file)}")
    speaker = interface.create_speaker(audio_file, whisper_model="large-v3-turbo")
    
    # Cache the speaker profile
    speaker_cache[cache_key] = speaker
    print(f"💾 Cached speaker profile ({len(speaker_cache)} total cached)")
    
    return speaker

def create_speaker_and_generate(model_name, audio_file, test_text: Optional[str] = None, temperature: float = 0.4):
    """Create speaker from audio and optionally generate test audio."""
    if audio_file is None:
        # Return default values for startup/caching purposes
        return "Please upload an audio file to create a speaker profile.", None
    
    # Get interface (no caching to avoid CUDA memory issues)
    interface = get_interface(model_name)
    
    # Get or create speaker profile (with caching)
    speaker = get_or_create_speaker(interface, audio_file)
    
    # Convert speaker dict to formatted JSON
    speaker_json = json.dumps(speaker, indent=2, ensure_ascii=False)
    
    # Generate test audio if text is provided
    generated_audio = None
    if test_text and test_text.strip():
        output = interface.generate(
            config=outetts.GenerationConfig(
                text=test_text,
                speaker=speaker,
                sampler_config=outetts.SamplerConfig(
                    temperature=temperature
                ),
                max_length=MODEL_INFO[MODELS[model_name]]["max_seq_length"]
            )
        )
        
        # Save to temporary file
        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
            output.save(f.name)
            generated_audio = f.name
    
    return speaker_json, generated_audio

example_text = "Hello, this is a test of the OuteTTS speaker profile."

# Create the Gradio interface
demo = gr.Interface(
    fn=create_speaker_and_generate,
    inputs=[
        gr.Dropdown(
            choices=list(MODELS.keys()),
            value=list(MODELS.keys())[-1],
            label="Select OuteTTS Model",
            info="Choose the model variant to use"
        ),
        gr.Audio(
            label="Upload Reference Audio",
            type="filepath",
            sources=["upload", "microphone"]
        ),
        gr.Textbox(
            label="Test Text (Optional)",
            placeholder="Enter text to generate speech (leave empty to only create speaker profile)...",
            lines=3,
            value=None
        ),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            step=0.1,
            value=0.4,
            label="Temperature",
            info="Controls randomness in generation"
        )
    ],
    outputs=[
        gr.Textbox(
            label="Speaker Profile (JSON)",
            lines=15,
            max_lines=20,
            show_copy_button=True
        ),
        gr.Audio(
            label="Generated Test Audio (if text provided)",
            type="filepath"
        )
    ],
    title="🎙️ OuteTTS Speaker Creator",
    description="Create and manage speaker profiles for OuteTTS text-to-speech synthesis. Upload audio to create a speaker profile, and optionally provide test text to generate sample audio.",
    theme=gr.themes.Soft(),
    examples=[
        ["OuteTTS-1.0-0.6B", None, example_text, 0.2],
        ["OuteTTS-0.3-500M", None, example_text, 0.2],
    ],
    cache_examples=False,
    flagging_mode="never"
)

if __name__ == "__main__":
    # Launch with optimized configuration for HuggingFace Spaces
    demo.launch(
        server_name="0.0.0.0",  # Allow external connections
        server_port=7860,
        share=False,           # Set to True if you want a public link
        show_api=True,         # Show API documentation
        show_error=True        # Show detailed error messages
    )