Dia
A 1.6B-parameter English text-to-speech model that generates dialogue with nonverbal sounds from a transcript.
Apache-2.0 model for multi-speaker English dialogue with laughs and coughs; CUDA GPU required.
Dia is a 1.6B-parameter text-to-speech model from Nari Labs that generates dialogue directly from a transcript. It suits developers who want expressive English dialogue on their own CUDA GPU.
What it is
You install it with pip, then run the example script, the web app or the command-line tool; weights load from Hugging Face. The README says it can condition on audio for emotion and tone, produce nonverbal sounds like laughter and coughing, and clone a voice if you supply the transcript of the audio.
Who it's for
It suits English-language podcast, game and prototype work for people with an Nvidia GPU. It does not suit multilingual projects or machines without CUDA, since CPU support is listed as coming.
Key features
Dialogue from transcript
Generates multi-speaker conversation from a script.
Nonverbal sounds
Laughter, coughing and throat clearing are supported.
Audio conditioning
Control emotion and tone from a reference clip.
Voice cloning
Needs the transcript of the audio you want to clone.
Precision options
bfloat16 uses about 4.4 GB VRAM; float32 about 7.9 GB.
Web app and CLI
Run app.py or cli.py after installing.
Up and running in 3 steps
Our honest verdict
What's great
- Apache-2.0 licence on the code and a documented pretrained 1.6B checkpoint.
- bfloat16 inference needs about 4.4 GB of VRAM on an RTX 4090 per the README.
- Handles nonverbal cues such as laughs, which many open TTS tools omit.
- Both a web app and a command-line tool ship in the repository.
Where it falls short
- English only at the moment, per the README.
- GPU only: the README lists CUDA 12.6 and says CPU support is still to come.
- RTX 5000-series cards need a torch 2.8 nightly build per the README.
- The README restricts identity misuse and illegal uses, and the model is described as for research and education.
Pricing & license
Free under Apache-2.0. The only cost is the CUDA GPU you run it on; the README mentions a waitlist for larger hosted versions but shows no price.
Open remote AI-training roles
768 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Does Dia run on CPU?
No. The README says it was only run on GPUs with CUDA 12.6 and that CPU support is to be added.
How much VRAM does Dia need?
The README lists about 4.4 GB for bfloat16 and float16 and about 7.9 GB for float32 on an RTX 4090.
Can Dia clone voices?
The README describes audio prompts for voice cloning and says to provide the transcript of the audio before your text.
