Dia2-1B

Dia2 is a streaming dialogue TTS model created by Nari Labs.

The model does not need the entire text to produce the audio, and can start generating as the first few words are given as input. You can condition the output on audio, enabling natural conversations in realtime.

We provide model checkpoints (1B, 2B) and inference code to accelerate research. The model only supports up to 2 minutes of generation in English.

Upcoming

• Dia2 TTS Server: Real streaming support
• Sori: Dia2-powered speech-to-speech engine written in Rust

Quickstart

Requirement — install uv and use CUDA 12.8+ drivers. All commands below run through uv run … as a rule.

1. Install dependencies (one-time):

   uv sync
   

2. Prepare a script: edit input.txt using [S1] / [S2] speaker tags.
3. Generate audio:

   uv run -m dia2.cli \
     --hf nari-labs/Dia2-1B \
     --input input.txt \
     --cfg 2.0 --temperature 0.8 \
     --cuda-graph --verbose \
     output.wav
   

   The first run downloads weights/tokenizer/Mimi. The CLI auto-selects CUDA when available (otherwise CPU) and defaults to bfloat16 precision—override with --device / --dtype if needed.
4. Conditional Generation (optional):

   uv run -m dia2.cli \
     --hf nari-labs/Dia2-1B \
     --input input.txt \
     --prefix-speaker-1 prefix_speaker1.wav \
     --prefix-speaker-2 prefix_speaker2.wav \
     --cuda-graph --verbose \
     output_conditioned.wav
   

   Condition the generation on previous conversational context in order to generate natural output for your speech-to-speech system. For example, place the voice of your assistant as prefix speaker 1, place user's audio input as prefix speaker 2, and generate the response to user's input.
5. Gradio for Easy Usage

   uv run gradio_app.py
   

Programmatic Usage

from dia2 import Dia2, GenerationConfig, SamplingConfig

dia = Dia2.from_repo("nari-labs/Dia2-1B", device="cuda", dtype="bfloat16")
config = GenerationConfig(
    cfg_scale=2.0,
    audio=SamplingConfig(temperature=0.8, top_k=50),
    use_cuda_graph=True,
)
result = dia.generate("[S1] Hello Dia2!", config=config, output_wav="hello.wav", verbose=True)

Generation runs until the runtime config's max_context_steps (1500, 2 minutes) or until EOS is detected. GenerationResult includes audio tokens, waveform tensor, and word timestamps relative to Mimi’s ~12.5 Hz frame rate.

Hugging Face

License & Attribution

Licensed under Apache 2.0. All third-party assets (Kyutai Mimi codec, etc.) retain their original licenses.

Disclaimer

This project offers a high-fidelity speech generation model intended for research and educational use. The following uses are strictly forbidden:

• Identity Misuse: Do not produce audio resembling real individuals without permission.
• Deceptive Content: Do not use this model to generate misleading content (e.g. fake news)
• Illegal or Malicious Use: Do not use this model for activities that are illegal or intended to cause harm.

By using this model, you agree to uphold relevant legal standards and ethical responsibilities. We are not responsible for any misuse and firmly oppose any unethical usage of this technology.

Acknowledgements

• We thank the TPU Research Cloud program for providing compute for training.
• Our work was heavily inspired by KyutaiTTS and Sesame

Questions? Join our Discord or open an issue.