🗣 Dia is a new open source text-to-speech model from Nari Labs with 1.6 billion parameters that can generate rich, full-fledged dialogue.

Key features:
- Ultra-realistic dialogue. Generation of coordinated lines of two "speaking" characters, tagged [S1] and [S2] in one text.

- Emotions and tone. You can set the tone and intonation through an acoustic request (audio prompt), and also control the "non-verbal": laughter, coughing, sighs, etc.

- Voice cloning. Voice cloning based on a short sample: upload audio and its transcript, and the model will adapt to the specified timbre
GitHub

The model is written in Python (100% of the code) using PyTorch 2.0 and CUDA 12.6

Performance and requirements:

The full version requires ≈10 GB of VRAM; quantization of the model is planned in the future.

Installation and launch:


pip install git+https://github.com/nari-labs/dia.git
git clone https://github.com/nari-labs/dia.git
cd dia
uv run app.py # or python app.py

The Gradio interface immediately shows the difference with ElevenLabs and Sesame CSM‑1B

License: Apache 2.0.

Dia is great for ML research in TTS: you get open weight files, a flexible scripting API, and a UI for quickly testing hypotheses.

Dia currently only supports speech generation in English.

▪️Demo
▪️ Github
▪️HF