OmniVoice
k2-fsa/OmniVoiceReviewed by AI
Checked by RepoPicks Radar on
What is OmniVoice?
OmniVoice is a text-to-speech model you run yourself. Give it 3 to 10 seconds of someone speaking and it reads your text in that voice, or describe a voice instead (female, low pitch, British accent) and it invents one. It can add laughter or sighs and fix pronunciations. It runs on an NVIDIA, Apple Silicon or Intel Arc graphics chip, with a simple web page started from the terminal; the authors also host a free online demo you can try without installing anything. Free and Apache-2.0, with a 600+ language claim from the authors' paper.
OmniVoice is free and open source (Apache-2.0) and can replace ElevenLabs and Descript.
How easy is it to install OmniVoice?
A · download and open · Web
- Windows D
- macOS D
- Linux D
- Web A
The easiest way in is the web: the authors keep a free Hugging Face Space (and a Google Colab notebook), so you open a page and type text, which is why web is A. Installing it is D on all three systems: you create a Python environment, install the right PyTorch build for your exact graphics chip (CUDA 12.8 wheels for NVIDIA, a separate Intel Arc index, Apple Silicon for Macs), then `pip install omnivoice` and run `omnivoice-demo` from the terminal.
OmniVoice on GitHub
- Stars
- 16,480on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +192stars in 6 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux, Web
- You need to install
- Python, NVIDIA graphics card
- To build from source
- Python, NVIDIA graphics card
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- None
- Cost
- Free · Free software and free model weights. The cost is your own hardware: a graphics card and roughly 1 to 4.5 GB of disk for the downloaded model.
- License
- Apache-2.0 (Permissive)
- Can replace
- ElevenLabs, Descript
- Last change on GitHub
- Aug 24, 2026
OmniVoice risk signals (6)
- Attentionvoice cloning can impersonate people
README.md Disclaimer: "Users are strictly prohibited from using this model for unauthorized voice cloning, voice impersonation, fraud, scams, or any other illegal or unethical activities." The tool clones a voice from a 3 to 10 second clip, so it can copy anyone whose recording you have, and nothing in the code checks consent.
AI review
- Attentionlocal web demo opens to the network
The documented command is `omnivoice-demo --ip 0.0.0.0 --port 8001` (README.md Quick Start and Command-Line Tools). Binding to 0.0.0.0 means every device on your network can open the voice-cloning page, and no password is mentioned.
AI review
- Infodownloads models on first run
`OmniVoice.from_pretrained("k2-fsa/OmniVoice")` fetches the weights from Hugging Face the first time, and when you omit `ref_text` it also downloads a Whisper model to transcribe your reference audio. The checks estimate 950 MB to 4450 MB of disk.
AI review
- Infothird party download mirrors suggested
README.md suggests `export HF_ENDPOINT="https://hf-mirror.com"` for model downloads and `uv sync --default-index "https://mirrors.aliyun.com/pypi/simple"` for packages. Both route the download through a mirror that is not the official source.
AI review
- Infoquality claims are the authors own
The "600+ languages", "state-of-the-art" and speed numbers come from the authors' own README and paper (arXiv:2604.00688). The README itself warns that voice design "is trained on Chinese and English data only" and "may produce unstable results for some low-resource languages", and that cross-language cloning carries the reference voice's accent.
AI review
- Infoearly version research code
Version 0.2.1 with six releases, the newest from 2026-07-16, and the deterministic checks record the last commit on 2026-08-24, 71 days before the check. pyproject.toml lists the audience as "Science/Research" and "Developers".
AI review
Evidence is shown in its original language.
What Radar could not check about OmniVoice
- How good the cloned voice sounds in your language, especially outside Chinese and English
- Whether it runs usefully on a computer without a graphics card, and how slow that is
- Real disk and memory use after the model and the Whisper transcriber are downloaded
- Whether the hosted online demo stays available and what it does with the audio you upload
- Whether the speed figures hold on anything other than the H100 card used in the benchmark