Verified project record

nari-labs/dia

Dia is a self-hosted text-to-audio dialogue model designed for research and educational use, generating multi-speaker conversations with non-verbal sounds and voice cloning from audio prompts. A CUDA-capable GPU is required, as CPU inference is not currently available.

Project overview

Project type: Model Runtime · Audio & Speech
Use cases: Audio & Speech
Deployment: Refer to project documentation
License: Apache-2.0

Best for

Researchers and educators exploring dialogue generation, emotion control, and voice cloning in a local self-hosted environment.

Key capabilities

Generate dialogue via `[S1]` and `[S2]` tags
Generate non-verbal sounds like (laughs), (coughs), etc.
Clone voices using audio prompts
Provides a web-based user interface via Gradio
Operate the model via a command-line interface
Planned Docker support for ARM architecture and MacOS
Planned addition of quantization for memory efficiency

Limitations and risks

Only supports English generation at the moment.
CPU support is to be added soon; the model has been tested only on GPUs.
Outputs differ on every run unless an audio prompt or seed is fixed, as the model is not fine-tuned on a specific voice.
Short input (<5s) sounds unnatural, and very long input (>20s) makes speech unnaturally fast.
There is a risk of identity misuse, deceptive content generation, and illegal or malicious use.

Getting started

Clone the repository, install dependencies using pip, and launch the web interface or command line interface. Setup difficulty is medium due to the Python, PyTorch, and CUDA requirements.

Evidence and sources

GitHub project description: A TTS model capable of generating ultra-realistic dialogue in one pass.
README: Dia directly generates highly realistic dialogue from a transcript. You can condition the output on audio, enabling emotion and tone control. The model can also produce nonverbal…
README: To accelerate research, we are providing access to pretrained model checkpoints and inference code. The model weights are hosted on Hugging Face.
README: This project offers a high-fidelity speech generation model intended for research and educational use.
README: Docker support for ARM architecture and MacOS.

Reviewed at: 2026-07-17T12:04:18Z