dia — AI 语音合成工具中文文档

or install with uv

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

Run `hf.py`. The file is as below.

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") ```

</details>

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.

</details>


<details>
<summary>Install via uv</summary>

You need [uv](https://docs.astral.sh/uv/) to be installed.

Or if you have uv installed

uv run app.py

</details>

<details>
<summary>Run with CLI</summary>

Or if you have uv installed

uv run cli.py --help ```

</details>

[!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.

Generation Guidelines

- Keep input text length moderate - Short input (corresponding to under 5s of audio) will sound unnatural - Very long input (corresponding to over 20s of audio) will make the speech unnaturally fast. - Use non-verbal tags sparingly, from the list in the README. Overusing or using unlisted non-verbals may cause weird artifacts. - Always begin input text with [S1], and always alternate between [S1] and [S2] (i.e. [S1]... [S1]... is not good) - When using audio prompts (voice cloning), follow these instructions carefully: - Provide the transcript of the to-be cloned audio before the generation text. - Transcript must use [S1], [S2] speaker tags correctly (i.e. single speaker: [S1]..., two speakers: [S1]... [S2]...) - Duration of the to-be cloned audio should be 5~10 seconds for the best results. (Keep in mind: 1 second ≈ 86 tokens) - Put [S1] or [S2] (the second-to-last speaker's tag) at the end of the audio to improve audio quality at the end

Quickstart

Optionally

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