Dia-0626

Dia is a 1.6B parameter text to speech model created by Nari Labs. It was pushed to the Hub using the PytorchModelHubMixin integration.

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.

• (Update) We have a ZeroGPU Space running! Try it now here. Thanks to the HF team for the support :)
• Play with a larger version of Dia: generate fun conversations, remix content, and share with friends. 🔮 Join the waitlist for early access.

⚡️ Quickstart

This will open a Gradio UI that you can work on.

git clone https://github.com/nari-labs/dia.git
cd dia && uv run app.py

or if you do not have uv pre-installed:

git clone https://github.com/nari-labs/dia.git
cd dia
python -m venv .venv
source .venv/bin/activate
pip install uv
uv run app.py

Note that 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 (a guide coming VERY soon - try it with the second example on Gradio for now), or fixing the seed.

Features

• Generate dialogue via [S1] and [S2] tag
• Generate non-verbal like (laughs), (coughs), etc.
  • Below verbal tags will be recognized, but might result in unexpected output.
  • (laughs), (clears throat), (sighs), (gasps), (coughs), (singing), (sings), (mumbles), (beep), (groans), (sniffs), (claps), (screams), (inhales), (exhales), (applause), (burps), (humming), (sneezes), (chuckle), (whistles)
• Voice cloning. See example/voice_clone.py for more information.
  • In the Hugging Face space, you can upload the audio you want to clone and place its transcript before your script. Make sure the transcript follows the required format. The model will then output only the content of your script.

⚙️ Usage

As a Python Library

import soundfile as sf

from dia.model import Dia

model = Dia.from_pretrained("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."

output = model.generate(text)

sf.write("simple.mp3", output, 44100)

A pypi package and a working CLI tool will be available soon.

As part of transformers

Install transformers:

# pip
pip install "transformers[torch]"

# uv
uv pip install "transformers[torch]"

Generation with Text

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."]
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=256)  # corresponds to around ~2s

# save audio to a file
outputs = processor.batch_decode(outputs)
processor.save_audio(outputs, "example.wav")

Generation with Text and Audio (Voice Cloning)

from datasets import load_dataset, Audio
from transformers import AutoProcessor, DiaForConditionalGeneration

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

ds = load_dataset("hf-internal-testing/dailytalk-dummy", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=44100))
audio = ds[-1]["audio"]["array"]
# text is a transcript of the audio + additional text you want as new audio
text = ["[S1] I know. It's going to save me a lot of money, I hope. [S2] I sure hope so for you."]

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

model = DiaForConditionalGeneration.from_pretrained(model_checkpoint).to(torch_device)
outputs = model.generate(**inputs, max_new_tokens=256)  # corresponds to around ~2s

# retrieve actually generated audio and save to a file
outputs = processor.batch_decode(outputs, audio_prompt_len=prompt_len)
processor.save_audio(outputs, "example_with_audio.wav")

💻 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.

On enterprise GPUs, Dia can generate audio in real-time. On older GPUs, inference time will be slower. For reference, on a A4000 GPU, Dia roughly generates 40 tokens/s (86 tokens equals 1 second of audio). torch.compile will increase speeds for supported GPUs.

The full version of Dia requires around 10GB of VRAM to run. 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