Mac Dia Server

FastAPI server providing an OpenAI-compatible Text-to-Speech (TTS) API endpoint, utilizing mlx-audio for generation on Apple Silicon.

About Dia-1.6B

Dia-1.6B is a state-of-the-art open-source text-to-speech (TTS) model developed by Nari Labs, featuring 1.6 billion parameters. This service is dedicated to providing TTS capabilities using the fixed model mlx-community/Dia-1.6B-4bit. Key features include:

• Multi-speaker dialogue generation using [S1], [S2] tags in the input text
• Fine-grained control over voice, emotion, and speaking style
• Support for non-verbal expressions like laughter, coughing, and more
• Voice cloning capabilities for personalized speech synthesis
• Optimized for English language generation

Dia-1.6B is comparable in performance to leading commercial TTS solutions, while remaining fully open and customizable for research and production use.

Setup

1. Install Dependencies: Requires Python 3.12+ and uv.

   uv venv  # Create virtual environment
   source .venv/bin/activate
   uv pip install .
   # Special attention might be needed for installing mlx and mlx-audio.
   # Follow official MLX documentation.
   uv run start.py

2. Configure API Key: Create a .env file in the project root:

   API_KEY=your_actual_api_key
   

3. Run the Server:

   uv run start.py

API Endpoint

• URL: /v1/audio/speech
• Method: POST
• Authentication: Authorization: Bearer <YOUR_API_KEY>
• Request Body: (See OpenAI TTS API documentation)
  • model (string): e.g., "tts-1"
  • input (string): Text to synthesize.
  • voice (string): e.g., "alloy"
  • response_format (string, optional): e.g., "mp3", defaults to "mp3".
  • speed (float, optional): Speed, defaults to 1.
• Response: Audio stream in the specified format.

CURl

  curl -X 'POST' \
  'http://localhost:8000/v1/audio/speech' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -H "Authorization: Bearer " \
  -d '{
  "model": "string",
  "input": "[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face.",
  "voice": "alloy",
  "response_format": "mp3",
  "speed": 1
 }' \
    --output speech.mp3

curl -X POST http://localhost:8000/v1/audio/transcriptions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F file=@/yourfile \
  -F model=mlx-community/whisper-large-v3-turbo \
  -F language=en