OmniVoice
Xiaomi's Next-gen Kaldi team TTS for 600+ languages with cloning and voice design. Code is Apache 2.0, but the weights are CC BY-NC.
In short. OmniVoice turns text into speech in more than 600 languages. It can copy a voice from a short recording or create a new voice from a description such as age, gender and accent. It is fast, but the downloadable model is licensed for non-commercial use only.
Summary
OmniVoice comes from the Next-gen Kaldi (k2-fsa) speech team at Xiaomi and was published with a paper on 1 April 2026. It is a 0.6B-parameter model, initialized from Qwen3-0.6B, that maps text straight to audio tokens with a diffusion-language-model-style, non-autoregressive design instead of generating one token after another. It was trained on a 581,000-hour multilingual dataset assembled entirely from open-source data and supports more than 600 languages, which its authors call the broadest coverage of any zero-shot TTS model.
It clones a voice from a short reference clip, designs voices from attributes such as gender, age, pitch and accent, and accepts tags such as [laughter] and pinyin or phoneme hints to fix pronunciation. Voice cloning is the most stable mode, and voice design was trained on Chinese and English only. The card reports a real-time factor as low as 0.025, about 40 times faster than real time, and the GitHub page measures 0.0115 on an H100 with FlashInfer at batch size 8. Output is 24 kHz audio.
The license has a catch. The code is Apache 2.0, but the pre-trained model is CC BY-NC because of constraints in its training data, such as Emilia, so you cannot use it commercially. The authors also prohibit unauthorized voice cloning and impersonation. The checkpoint is a 3.3 GB float32 file, and the omnivoice Python package installs the runtime.
Good for
- Text-to-speech in low-resource languages that other models do not cover
- Research and personal projects that need voice cloning or voice design
- Fast batch generation on a single GPU
Watch out for
- The weights are CC BY-NC, so commercial use is not allowed, even though the code is Apache 2.0
- Voice design was trained on Chinese and English only, and results can be unstable in low-resource languages
- Cross-language cloning carries the accent of the reference clip
- The authors prohibit unauthorized voice cloning, impersonation and fraud
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of OmniVoice. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| OmniVoice FP32 | 613M | 3.26 GB | 3.75 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
A Creative Commons license that lets anyone copy, share and adapt the work with credit, but only for non-commercial purposes. Some model cards name it without a version number, and the current version is 4.0.
- Attribution or model name must be shown
See also
- OmniVoice vs VoxCPM2: which open text-to-speech model should you use?
- Best open-source speech models for transcription and text-to-speech
- Best open-source text-to-speech models in 2026
- Open speech models: transcription is open, text-to-speech is mixed
Frequently asked questions
- What is OmniVoice?
- OmniVoice turns text into speech in more than 600 languages. It can copy a voice from a short recording or create a new voice from a description such as age, gender and accent. It is fast, but the downloadable model is licensed for non-commercial use only.
- Can I use OmniVoice commercially?
- Not without separate permission. OmniVoice is released under the CC BY-NC, which allows research and personal use only.
- Is OmniVoice open source?
- OmniVoice is open-weight, not open source in the strict sense. You can download the model files, but the CC BY-NC adds conditions that the Open Source Initiative definition does not allow.
- Can I run OmniVoice on my own computer?
- Yes. The smallest variant, OmniVoice, needs about 3.75 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was OmniVoice released?
- OmniVoice was released by Xiaomi on April 1, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models (arXiv, 1 Apr 2026) arxiv.org
- OmniVoice model card (Hugging Face) huggingface.co
- OmniVoice repository (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| VoxCPM2 OpenBMB | 2.29B | – | Apache 2.0 (Permissive) | 8 GB Laptop or phone | |
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone | |
| Chatterbox Resemble AI | 500M | – | MIT (Permissive) | 3.69 GB Laptop or phone | |
| LFM3-Audio-2B Liquid AI | 2B | 32K | LFM Open 1.0 (Conditional) | 4.6 GB Laptop or phone |