Dia
nari-labs/diaReviewed by AI
Checked by RepoPicks Radar on
What is Dia?
Dia is a speech model from Nari Labs that reads a script written with [S1] and [S2] speaker tags and produces a conversation that sounds like two people talking, including sounds like laughter, coughing or sighing. You can hand it 5 to 10 seconds of recorded audio to copy a voice and keep it steady. English only. There is a free page on Hugging Face where you can try it; to run it yourself you need Python and an NVIDIA card with about 4.4 GB of video memory, since the README says the processor alone is not supported yet.
Dia is free and open source (Apache-2.0) and can replace ElevenLabs, Murf, and NotebookLM.
How easy is it to install Dia?
A · download and open · Web
- Windows D
- Linux D
- Web A
The easy way is the maintainers' free Hugging Face page ("We have a ZeroGPU Space running"), which needs nothing installed. Running it yourself on Windows or Linux means cloning the repository, installing Python packages pinned to torch 2.6 with CUDA 12.6, and an NVIDIA card: "Dia has been tested on only GPUs (pytorch 2.0+, CUDA 12.6). CPU support is to be added soon." macOS support is still on the TODO list.
Dia on GitHub
- Stars
- 19,411on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +3stars in 6 days
Stars over time
Facts
- Runs on
- Windows, Linux, Web
- You need to install
- Nothing extra
- To build from source
- Python, NVIDIA graphics card
- Disk space
- 900 MB–4.2 GB · estimate
- Accounts or keys
- None
- Cost
- Free · Code and weights are free (Apache 2.0, model files on Hugging Face). The README invites you to a waitlist for "a larger version of Dia", whose terms are not stated.
- License
- Apache-2.0 (Permissive)
- Can replace
- ElevenLabs, Murf, NotebookLM
- Last change on GitHub
- Nov 19, 2025
Dia risk signals (4)
- Attentionvoice cloning impersonation risk
The Features list includes "Voice cloning" from 5-10 seconds of audio, and the project's own disclaimer forbids "**Identity Misuse**: Do not produce audio resembling real individuals without permission" and "**Deceptive Content**" such as fake news.
AI review
- Attentionsuperseded by newer repository
README: "**UPDATE 🚀(11/19)**: Dia2 is released on Github and HuggingFace", and the last commit here is 2025-11-19, 354 days before this check, with 91 issues open.
AI review
- Infonvidia only no cpu
"Dia has been tested on only GPUs (pytorch 2.0+, CUDA 12.6). CPU support is to be added soon", and the TODO list still has "Docker support for ARM architecture and MacOS".
AI review
- Infoenglish only and random voices
"The model only supports English generation at the moment" and "The model was not fine-tuned on a specific voice. Hence, you will get different voices every time you run the model" unless you fix the seed or give an audio sample.
AI review
Evidence is shown in its original language.
What Radar could not check about Dia
- How good the result sounds outside the maintainers' own demo page
- Total disk space after the model and the Descript Audio Codec download on first run
- Whether it works on cards other than the RTX 4090 the speed table used, including the 5000 series note
- Whether the free Hugging Face page stays available and what it keeps of the audio you upload
- Whether the repository will get fixes now that work moved to the Dia2 repository