dia README

The project's README from github.com/nari-labs/dia. As one markdown file: README.md.

Static Badge LICENSE

Model on HuggingFace Space on HuggingFace

Dia is a 1.6B parameter text to speech model created by Nari Labs.

UPDATE 🤗(06/27): Dia is now available through Hugging Face Transformers!

UPDATE 🚀(11/19): Dia2 is released on Github and HuggingFace link!

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 communications like laughter, coughing, clearing throat, etc.

To accelerate research, we are providing access to pretrained model checkpoints and inference code. The model weights are hosted on Hugging Face. The model only supports English generation at the moment.

We also provide a demo page comparing our model to ElevenLabs Studio and Sesame CSM-1B.

Generation Guidelines

Quickstart

Transformers Support

We now have a Hugging Face Transformers implementation of Dia! You should install main branch of transformers to use it. See hf.py for more information.

View more details

Install main branch of transformers

pip install git+https://github.com/huggingface/transformers.git
# or install with uv
uv pip install git+https://github.com/huggingface/transformers.git

Run hf.py. The file is as below.

from transformers import AutoProcessor, DiaForConditionalGeneration


torch_device = "cuda"
model_checkpoint = "nari-labs/Dia-1.6B-0626"

text = [
    "[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."
]
processor = AutoProcessor.from_pretrained(model_checkpoint)
inputs = processor(text=text, padding=True, return_tensors="pt").to(torch_device)

model = DiaForConditionalGeneration.from_pretrained(model_checkpoint).to(torch_device)
outputs = model.generate(
    **inputs, max_new_tokens=3072, guidance_scale=3.0, temperature=1.8, top_p=0.90, top_k=45
)

outputs = processor.batch_decode(outputs)
processor.save_audio(outputs, "example.mp3")

Run with this repo

Install via pip 
# Clone this repository
git clone https://github.com/nari-labs/dia.git
cd dia

# Optionally
python -m venv .venv && source .venv/bin/activate

# Install dia
pip install -e .

Or you can install without cloning.

# Install directly from GitHub
pip install git+https://github.com/nari-labs/dia.git

Now, run some examples.

python example/simple.py
Install via uv

You need uv to be installed.

# Clone this repository
git clone https://github.com/nari-labs/dia.git
cd dia

Run some examples directly.

uv run example/simple.py
Run Gradio UI 
python app.py

# Or if you have uv installed
uv run app.py
Run with CLI 
python cli.py --help

# Or if you have uv installed
uv run cli.py --help

[!NOTE] The model was not fine-tuned on a specific voice. Hence, you will get different voices every time you run the model. You can keep speaker consistency by either adding an audio prompt, or fixing the seed.

[!IMPORTANT] If you are using 5000 series GPU, you should use torch 2.8 nightly. Look at the issue #26 for more details.

Features

💻 Hardware and Inference Speed

Dia has been tested on only GPUs (pytorch 2.0+, CUDA 12.6). CPU support is to be added soon. The initial run will take longer as the Descript Audio Codec also needs to be downloaded.

These are the speed we benchmarked in RTX 4090.

precision realtime factor w/ compile realtime factor w/o compile VRAM
bfloat16 x2.1 x1.5 ~4.4GB
float16 x2.2 x1.3 ~4.4GB
float32 x1 x0.9 ~7.9GB

We will be adding a quantized version in the future.

If you don't have hardware available or if you want to play with bigger versions of our models, join the waitlist here.

🪪 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

⚠️ Disclaimer

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

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.

🔭 TODO / Future Work

🤝 Contributing

We are a tiny team of 1 full-time and 1 part-time research-engineers. We are extra-welcome to any contributions! Join our Discord Server for discussions.

🤗 Acknowledgements

⭐ Star History

Star History Chart

dia (nari-labs/dia)

Searching Database…

19.4K ·GitHub
Dia
What is this?Dia is a 1.6B-parameter open-weights text-to-speech model from Nari Labs that generates ultra-realistic multi-speaker dialogue from a transcript in a single pass.
Is this a web service or something I run myself?You run it yourself: it is a Python model with inference code, and the weights are downloaded from Hugging Face. If you just want to try it first, there is a free ZeroGPU Space that runs in the browser.
How do I get it running?Install it with pip or uv and run the example scripts. There is also a Hugging Face Transformers implementation (install the main branch of transformers, then run hf.py), plus a Gradio UI and a CLI.
What are the main limitations to know upfront?Generation is English-only for now. Keep clips in the 5-20 second range: shorter input sounds unnatural, longer input makes the speech unnaturally fast. RTX 5000-series cards currently need torch 2.8 nightly.
Is it free to use, and are there usage restrictions?It is Apache 2.0 licensed, but the disclaimer restricts use to research and education: no cloning real people's voices without permission, no deceptive content like fake news, and nothing illegal.
What kind of hardware does it need?A CUDA GPU — it has been tested on GPUs only (PyTorch 2.0+, CUDA 12.6), and CPU support is not in yet. In bfloat16 it needs about 4.4GB of VRAM and runs roughly 2x realtime on an RTX 4090.
How does it compare to ElevenLabs or other TTS models?The team publishes a demo page comparing Dia against ElevenLabs Studio and Sesame CSM-1B so you can judge quality by ear yourself; the README draws no verdict of its own.
Is Dia still actively developed?Yes — Dia was integrated into Hugging Face Transformers in the June update, and Dia2 was released on GitHub and Hugging Face in the November update.
What is on the roadmap?Docker support for ARM architecture and MacOS, faster inference, and a quantized version for lower memory use. There is also a waitlist for larger Dia versions and for people without the hardware to run it.
What are similar projects?csm, bark, ChatTTS, Speech, DramaBox, FireRedAudio, breeze-tts, OpenVoice
Which topics is it in?audio-generation, english, voice-cloning, speech-synthesis, voice, tts, text-to-speech, audio, inference, generative-ai, huggingface, cuda
Creator
nari-labs
Size
0.6 MB
Stars
19.4K
Updated
CommitNov 19
Age
1y
Commits · 90 Days