Dia

Dia is an open-source 1.6B parameter text-to-speech model developed by Nari Labs, designed to generate highly realistic dialogue directly from transcripts. It supports multi-speaker audio generation, non-verbal cues like laughter and coughing, and fine-grained emotion and tone control. Dia can also perform voice cloning using an audio reference, making it a powerful tool for content creators, researchers, and developers building conversational AI applications.

• Multi-speaker dialogue generation: Generate realistic conversations between multiple speakers directly from a text transcript using speaker tags.
• Non-verbal audio support: Include sounds like laughter, coughing, and sighs in generated audio by adding special tokens in the transcript.
• Emotion and tone control: Guide the emotional delivery of speech through natural language descriptions embedded in the transcript.
• Voice cloning: Provide an audio reference clip to clone a specific voice and use it in generated dialogue.
• Open-source model weights: Download and run the 1.6B parameter model locally via Hugging Face or the GitHub repository.
• Gradio demo: Try Dia instantly through the hosted Hugging Face Spaces demo without any local setup.
• Python API: Integrate Dia into your own applications using the provided Python package and inference scripts.
• Local inference: Run the model on your own hardware for full control over privacy and customization.