An OpenAI-compatible API for Dia text-to-speech, providing a drop-in replacement for applications using OpenAI's TTS service.

Features

• 🎯 OpenAI API Compatibility: Implements the OpenAI /audio/speech endpoint for seamless integration with existing applications.
• 🔊 Dia 1.6B Model: Uses Nari Lab's Dia 1.6B text-to-speech model for high-quality voice synthesis.
• 🎭 Voice Customization: Create and manage custom voices by uploading audio samples for voice cloning.
• 🗣 Multi-speaker Support: Use the [S1] and [S2] tags familiar to Dia users for multi-speaker dialogues.
• 📊 Optimized Performance: GPU acceleration with CUDA support for faster inference.
• 🐳 Docker Ready: Easy deployment using Docker with NVIDIA GPU support.

Quick Start

Prerequisites

• Docker and Docker Compose
• NVIDIA GPU with CUDA support (recommended)
• NVIDIA Container Toolkit for GPU acceleration

Installation & Startup

1. Clone the repository:

git clone https://github.com/phildougherty/dia_openai.git
cd dia_openai

2. Start the service with Docker Compose:

docker-compose up -d

3. Access the API:

The API will be available at: http://localhost:8000

API documentation: http://localhost:8000/docs

API Usage

Text-to-Speech (OpenAI Compatible)

import requests
import json

url = "http://localhost:8000/v1/audio/speech"

payload = {
    "model": "dia-1.6b",
    "input": "[S1] Hello, I'm speaking with the Dia TTS model. [S2] And I'm responding with a different voice.",
    "voice": "alloy",  # Standard voices: alloy, echo, fable, onyx, nova, shimmer
    "response_format": "mp3"
}

headers = {
    "Content-Type": "application/json"
}

response = requests.post(url, json=payload, headers=headers)

# Save the audio file
if response.status_code == 200:
    with open("output.mp3", "wb") as file:
        file.write(response.content)
    print("Audio saved as output.mp3")
else:
    print(f"Error: {response.status_code}")
    print(response.text)

Working with Custom Voices

Creating a Custom Voice

import requests

url = "http://localhost:8000/v1/audio/voices"

# Prepare multipart form data
files = {
    'file': ('sample.wav', open('path/to/voice_sample.wav', 'rb'), 'audio/wav')
}
data = {
    'name': 'My Custom Voice',
    'description': 'A custom voice created from my audio sample'
}

response = requests.post(url, files=files, data=data)

if response.status_code == 201:
    voice_data = response.json()
    voice_id = voice_data['voice_id']
    print(f"Created custom voice with ID: {voice_id}")
else:
    print(f"Error: {response.status_code}")
    print(response.text)

Using a Custom Voice

import requests

url = "http://localhost:8000/v1/audio/speech"

payload = {
    "model": "dia-1.6b",
    "input": "This is my custom voice speaking through the Dia model.",
    "voice": "custom_abc123def", # Replace with your actual custom voice ID
    "response_format": "mp3"
}

headers = {
    "Content-Type": "application/json"
}

response = requests.post(url, json=payload, headers=headers)

# Save the audio file
if response.status_code == 200:
    with open("custom_voice_output.mp3", "wb") as file:
        file.write(response.content)
    print("Custom voice audio saved")
else:
    print(f"Error: {response.status_code}")
    print(response.text)

Configuration

The application can be configured using environment variables in the docker-compose.yml file:

environment:
  - HOST=0.0.0.0
  - PORT=8000
  - DEBUG=false  # Set to true for more logs
  - ENABLE_CORS=true
  - USE_TORCH_COMPILE=true  # Set to false for better compatibility on some systems
  - COMPUTE_DTYPE=float16  # Options: float16, bfloat16, float32
  - OUTPUT_FORMAT=mp3
  - MAX_AUDIO_LENGTH_SEC=60

Advanced Configuration

Customizing Model Paths

If you have a custom model version, you can mount it by modifying the docker-compose.yml:

volumes:
  - ./my_custom_model:/app/models/dia-1.6b

Performance Tuning

For better performance on GPUs with limited VRAM, try the following settings:

environment:
  - COMPUTE_DTYPE=float16
  - USE_TORCH_COMPILE=false

For maximum quality but higher VRAM usage:

environment:
  - COMPUTE_DTYPE=float32
  - USE_TORCH_COMPILE=true

Project Structure

dia_openai/
├── app/                # Main application code
│   ├── api/            # API routes and schemas
│   ├── core/           # Core configuration
│   ├── models/         # Data models
│   ├── services/       # Business logic services
│   └── utils/          # Helper utilities
├── cache/              # Runtime cache directory
├── voices/             # Custom voice metadata storage
├── static/             # Static files and audio samples
├── tests/              # Test cases
├── docker-compose.yml  # Docker Compose configuration
└── Dockerfile          # Docker build instructions

Troubleshooting

Common Issues

• CUDA out of memory: Try lowering MAX_AUDIO_LENGTH_SEC or using COMPUTE_DTYPE=float16
• Slow first-time speech generation: The DAC model is being downloaded; subsequent calls will be faster
• Audio quality issues: Try increasing the cfg_scale parameter in API calls for better quality

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Nari Labs for creating the Dia model
• OpenAI for the API design this project emulates

---

Note: This is an unofficial project and is not affiliated with Nari Labs or OpenAI.# dia_openai