What is OmniVoice?
OmniVoice is an open text-to-speech model from the k2-fsa project (Next-gen Kaldi), published on Hugging Face at the end of March 2026 with a paper on arXiv on 1 April. It clones a voice from a 3 to 10 second reference clip, designs a voice from attributes such as gender, age, pitch, accent or whisper, and accepts non-verbal tags like [laughter] and phoneme or pinyin corrections.
Its claim is coverage: more than 600 languages, trained on 581,000 hours of audio curated from open-source data. Architecturally it is a single-stage, non-autoregressive model in the style of a diffusion language model, initialised from Qwen3-0.6B, that maps text straight to acoustic tokens. The main checkpoint has about 613 million parameters and outputs 24 kHz audio.
The code is Apache 2.0, but the bundled audio tokenizer carries Boson AI's Higgs Audio 2 Community License. The latest package release is 0.2.1 (16 July 2026), and the weights have about 2.3 million downloads.
Is OmniVoice worth it?
OmniVoice's case rests on breadth: over 600 languages in one zero-shot model, where most open rivals such as Chatterbox, Kokoro and Qwen3-TTS cover a few dozen at most. For a product that has to speak Yoruba, Khmer or a regional Arabic dialect, it is one of very few open options.
It is also fast. The project reports a real-time factor as low as 0.025, about 40 times faster than real time, and its own H100 benchmark on the Seed-TTS Chinese test set shows 0.0899 at batch size 1, falling to 0.0115 at batch size 8 with FlashInfer. At roughly 613 million parameters it runs on CUDA, Apple Silicon and Intel Arc, and the maintainers report testing on a 4 GB Arc A310.
The licence is the catch. The model code is Apache 2.0, but the audio tokenizer shipped with the weights falls under Boson AI's Higgs Audio 2 Community License, which is built on the Llama 3 licence. It requires a 'Built with Higgs Materials' notice, and any service with more than 100,000 annual active users must request an expanded licence that Boson may refuse. That is a much lower ceiling than Llama's 700 million.
Quality claims beyond speed and coverage are the authors' own. The paper reports state-of-the-art results on Chinese, English and multilingual benchmarks, but we found no independent leaderboard entry, and the README warns that cross-lingual cloning carries the reference speaker's accent and that digits read better once normalised to words.
Where it earns its place
An NGO building a health-information hotline for rural East Africa
Read out scripted messages in local languages with a consistent voice
Clone one narrator from a short clip and generate prompts offline on a single GPU; the user base sits far below the 100,000 annual active user threshold, so the community licence covers it.
A localisation engineer at an e-learning company
Draft voice-overs in 20 languages before hiring native narrators
Batch-generate with the omnivoice CLI on an H100 at a real-time factor near 0.01; costs are GPU hours, and the drafts let reviewers catch script problems early.
A consumer app with two million monthly users
Add multilingual voice replies
Above 100,000 annual active users the Higgs Audio 2 licence requires Boson's permission, so either negotiate an expanded licence first or choose an Apache or MIT-licensed model.
Before you commit
- Setup
- Install PyTorch for your accelerator, then pip install omnivoice; omnivoice-demo starts a local web UI, and a Hugging Face Space and Colab notebook need no install.
- Integrations
- Python API and command-line tools; Hugging Face weights; Whisper is used to auto-transcribe the reference clip if you omit its text; reusable voice-clone prompts can be saved to disk.
- Data and privacy
- Self-hosted, so audio stays on your hardware. Code is Apache 2.0; the audio tokenizer's Higgs Audio 2 Community License requires attribution and an expanded licence above 100,000 annual active users. The README prohibits unauthorised cloning and impersonation.
- Cost at scale
- No API fees. At the authors' batch-8 figure of 0.0115 real-time factor on one H100, an hour of speech takes under a minute of GPU time.
- Limits
- Reference clips longer than 10 seconds slow inference and may hurt quality; cross-lingual cloning keeps an accent; flash attention is unavailable on Intel XPU; no independent leaderboard results found.
What's good
- More than 600 languages in one model
- Voice cloning from 3 to 10 seconds and voice design from attributes
- Real-time factor as low as 0.025 by the authors' figures
- Runs on CUDA, Apple Silicon and Intel Arc
What's not
- Tokenizer licence caps free use at 100,000 annual active users
- Quality claims are self-reported
- Cross-lingual cloning carries the reference accent
- No hosted API from the authors
Best for
Developers who need open, self-hosted speech in languages the commercial and open leaders do not cover.
OmniVoice alternatives
OmniVoice: common questions
- Is OmniVoice worth it?
- OmniVoice is the open TTS model to reach for when the language is not on anyone else's list, provided your lawyer reads the tokenizer licence. We rate it 4.0 out of 5.
- How much does OmniVoice cost?
- Free weights; code under Apache 2.0, audio tokenizer under the Boson Higgs Audio 2 Community License (expanded licence needed above 100,000 annual active users).
- What are the downsides of OmniVoice?
- Tokenizer licence caps free use at 100,000 annual active users; quality claims are self-reported; cross-lingual cloning carries the reference accent; no hosted API from the authors.
- Who is OmniVoice best for?
- Developers who need open, self-hosted speech in languages the commercial and open leaders do not cover.
- How much does OmniVoice cost for a team or in production?
- No API fees. At the authors' batch-8 figure of 0.0115 real-time factor on one H100, an hour of speech takes under a minute of GPU time.
- Is OmniVoice safe for company data?
- Self-hosted, so audio stays on your hardware. Code is Apache 2.0; the audio tokenizer's Higgs Audio 2 Community License requires attribution and an expanded licence above 100,000 annual active users. The README prohibits unauthorised cloning and impersonation.
- What are the best OmniVoice alternatives?
- The closest alternatives we have reviewed are Chatterbox, Qwen3-TTS, Kokoro-82M and Eleven v4 and v4 Turbo.
Answers checked Oct 26, 2026.
Make OmniVoice? Show this rating on your site
