diajax 0.0.2a1

Dia-JAX: A JAX port of Dia, the text-to-speech model for generating realistic dialogue from text with emotion and tone control.

Dia-JAX

An experimental JAX port of Dia, the 1.6B parameter text-to-speech model from Nari Labs

Dia-JAX is a work-in-progress port of the original PyTorch-based Dia model to JAX via Flax NNX.

Features

Just like the original Dia model, Dia-JAX aims to offer:

• Generate dialogue via [S1] and [S2] tags
• Generate non-verbal elements like (laughs), (coughs), etc.
  • Supported verbal tags: (laughs), (clears throat), (sighs), (gasps), (coughs), (singing), (sings), (mumbles), (beep), (groans), (sniffs), (claps), (screams), (inhales), (exhales), (applause), (burps), (humming), (sneezes), (chuckle), (whistles)
• Voice cloning with reference audio (TODO: currently not implemented)
• Quality comparable to commercial solutions like ElevenLabs Studio

Quickstart

Install via pip

pip install diajax

⚙️ Usage

Note: Currently only recommended for experimental/development use due to memory issues

Run from Command Line

# Generate audio with default settings
dia --text "[S1] Dear Jacks, to generate audio from text from any machine. (applause) [S2] Really? How! (screams) [S1] With flakes and an axe. (chuckles)"

# Or with custom parameters
dia --temperature 0.7 --cfg-filter-top-k 42

As a Python Library

import diajax
model, config = diajax.load()
output = diajax.generate(model, config, text)

import soundfile as sf
sf.write('dia.mp3', output, 44100)

Acknowledgments

This project is a port of the original Dia model by Nari Labs. We thank them for releasing their model and code, which made this port possible.

License

This project is licensed under the same terms as the original Dia model. See LICENSE for details.