OmniVoice — Voice Cloning TTS in 600+ Languages
k2-fsa's zero-shot multilingual model, embedded here from its own Hugging Face Space
Hundreds of languages out of one model, cloning from a few seconds of consented reference audio, and — when you have no recording at all — a speaker assembled from attributes like age, pitch, accent and whisper. All of it is testable below before anything gets installed.
The panel above is OmniVoice's own Hugging Face Space, from k2-fsa. Give the first run extra time while ZeroGPU comes up. The write-up around it is ours; the demo is theirs.
What OmniVoice is trying to do
Most open TTS models ship a set of voices and a handful of languages. The k2-fsa project went the other way: breadth first, with what it claims is the widest language coverage of any open model, plus two routes to a speaker — clone one from a few seconds of audio, or specify one attribute by attribute and let the model construct it.
The base is Qwen3-0.6B, wrapped in a diffusion language-model-style architecture aimed at speed without a quality penalty; the model card puts the real-time factor as low as 0.025 — roughly 40× faster than real time. It ships broadly, too: model and Space on Hugging Face, an arXiv paper, a GitHub repo, a Colab notebook, and omnivoice on pip.
Came here looking for the demo? It is directly above. Type something, choose a language, add a clip if you want it cloned, adjust the generation settings, and 24 kHz audio plays back in the browser.
Where it pulls ahead
Breadth of language, cloning, and speaker design — usually three separate tools — inside one open model.
Language coverage as the headline
The model card enumerates over 600 language and dialect codes — coverage is the thing this model is built around.
Zero-shot voice cloning
Hand it a short recording and new text comes back in that speaker's style. A clean clip helps; adding the clip's transcript helps more.
A speaker built from description
With no sample to work from, specify the speaker instead — gender, age, pitch, whisper style, English accent or Chinese dialect.
Faster than real time
The diffusion language-model-style architecture is where the speed comes from: a real-time factor down to 0.025 on the model card, about 40× faster than playback.
Open, and built on Qwen3
Qwen3-0.6B underneath, released as a Hugging Face model and Space, a pip package, and a public GitHub repository.
Rules that come with it
The project's own guidance rules out unauthorised cloning, impersonation and unlawful use. Only clone a voice when you have the speaker's permission.
From text to audio in four steps
Everything below happens in the hosted Space; nothing needs installing locally.
Pick which mode you need
In the demo above: Voice Clone if you have a recording to work from, Voice Design if you are describing a speaker instead.
Give it text and a language
Paste your script. Leave detection on, or pick the language yourself from the dropdown — worth doing when the text is short.
Supply the voice
Cloning: a short clean sample, plus its transcript if you have one. Designing: dial in age, pitch, accent, whisper style.
Adjust, generate, listen
Speed, duration, inference steps, guidance and denoise are all exposed. Set them, generate, and play back the 24 kHz result.
The specification, condensed
Pulled from the model card, the Space and the demo interface, and laid out so you can decide quickly.
What it is good for
It earns its place on projects that need many languages *and* control over who appears to be speaking.
Localisation and dubbing
Draft voice tracks in a dozen languages for product video, course material, narration, and creative work that has to travel.
Prototyping voice agents
Settle on personality, accent and pacing before anyone builds the conversational or phone-agent stack around it.
Audio alternatives
Articles, docs and course material spoken aloud, for people who would rather listen or who need to.
Exploring characters and styles
Sweep through age, pitch, whisper and accent options via voice design instead of finding a reference speaker for each variant.
Benchmarking against your own stack
Put its zero-shot cloning, multilingual pronunciation and attribute control up against whatever TTS baseline you already trust.
Getting hands-on
pip install omnivoice, or clone the GitHub repo, and try Python inference, batch runs and your own pipeline around it.
Getting better takes
- One speaker, quiet room, normal speaking volume — that is what a good reference clip looks like.
- Make the first test short: pronunciation, pacing and speaker likeness are all judgeable in one sentence.
- Short or code-switched text is where detection slips — name the language yourself in those cases.
- Without consent from a real speaker, build the persona with voice design rather than cloning anyone.
- Drop in cues such as [laughter], and fix stubborn words with pinyin or phonemes, to tighten a take.
- Confidential scripts do not belong in a public Space — install the omnivoice package and run those locally.
OmniVoice, briefly answered
For anyone weighing it up for cloning, multilingual synthesis, or designed voices.
What OmniVoice is trying to do
A zero-shot TTS model from the k2-fsa project, built for breadth: 600+ languages, cloning from a short clip, and speakers designed from attributes when no clip exists.
Can I try OmniVoice online for free?
Yes — the official Hugging Face Space is embedded on this page, so the model runs in your browser before you install anything.
Does OmniVoice support voice cloning?
It does. Voice Clone mode takes a short reference clip, plus its transcript if you supply one, and generates in that voice. Restrict this to voices you own or have been given explicit permission to use.
What is OmniVoice voice design?
It is the no-reference path: describe the speaker — gender, age, pitch, whisper style, English accent, Chinese dialect — and the model constructs a voice to match.
How many languages does OmniVoice support?
Over 600 language and dialect codes are listed on the model card, which is about as broad as open zero-shot TTS gets. Coverage is not uniform quality, though — try your specific language in the Space before you build on it.
Is OmniVoice open source?
Yes: Qwen3-0.6B underneath, Apache-2.0 on top, with the model, a Space, a GitHub repo and a pip package all public. Read the licence terms yourself before commercial use.
Is the hosted OmniVoice demo private?
No — that Space runs on Hugging Face, not on our infrastructure. Treat it as public. For anything confidential, install the omnivoice package and keep the job on your own machine.
Test it before you install anything
Synthesise in another language, try cloning with a voice you are allowed to use, and see how far the design controls go — all in the official demo.