OmniVoice is a zero-shot text-to-speech (TTS) model released by k2-fsa, the open-source speech technology project known as Next-gen Kaldi. Built on a diffusion language model, it supports more than 600 languages. A single model handles both “voice cloning,” which reproduces a speaker’s voice from a few seconds of sample audio, and “voice design,” which creates a new voice from attributes you describe in words, such as gender and age. The source code and trained models are published under the Apache 2.0 license, and the project was released alongside its paper in April 2026.
Key Features
- Support for 600+ languages: The project describes its coverage as the broadest among zero-shot TTS models. It was trained on roughly 581,000 hours of open-source multilingual speech data, covering not only major languages but also languages with fewer speakers
- Voice cloning from a few seconds of audio: Provide a short reference clip and it generates speech that reproduces that speaker’s voice. No per-speaker fine-tuning is required
- Voice design by attributes: Specify gender, age, pitch, accent, dialect, whispering and more to create a new voice without any reference audio
- Non-verbal expressions and pronunciation correction: Insert symbols such as
[laughter]or[sigh]into the text to add non-verbal expressions. You can also set pronunciation directly using pinyin for Chinese and CMU dictionary notation for English - Fast inference: Its design converts text directly into acoustic tokens, and the project cites an RTF as low as 0.025 (about 40x faster than real time). This is the project’s own measurement and will vary by environment
- Runs on a wide range of hardware: It supports NVIDIA GPUs (CUDA), Apple Silicon (MPS) and Intel Arc GPUs (XPU), and can fall back to CPU when no GPU is available. It installs as the Python package
omnivoiceviapip install omnivoice
Pricing
| Plan | Price | Key Features |
|---|---|---|
| Open source (self-hosted) | Free | Source code and trained models published under Apache 2.0. Commercial use permitted |
| Online demo | Free | Try it in the browser on Hugging Face Spaces |
Pricing is as of September 2026. We could not find a paid hosted API or SaaS version. See the official repository for the latest information.
The software itself is free, but if you run it on your own PC or in the cloud, the cost of compute such as GPUs is yours to bear.
Pros & Cons
✅ Pros
- You can produce narration audio not only in Japanese or English but also in languages that commercial services cover thinly, which widens the range of multilingual content you can create
- There are no usage fees and the license permits commercial use, so costs stay contained even when you generate audio in bulk
- Everything can run in your own environment, so you don’t have to send scripts or reference audio to an external service
- You can design a voice in words without reference audio, making it easy to prototype character voices or narrator roles
⚠️ Cons
- There is no official managed service or hosted API, so you need to set up and run the Python environment yourself
- As a community-run project, it has no enterprise support desk or uptime guarantee
- Although it supports 600+ languages, the amount of training data differs by language, so quality is not necessarily uniform
- Use for unauthorized voice cloning, impersonation, fraud and similar purposes is explicitly prohibited, and handling other people’s voices requires ethical care such as obtaining their consent
Comparison with Similar Services
| Service | Difference from OmniVoice |
|---|---|
| XTTS (Coqui) | Also supports zero-shot voice cloning, but covers only 17 languages |
| F5-TTS | An open-source multilingual TTS, but supports fewer languages than OmniVoice |
| OpenVoice (MyShell) | Strong at converting voice timbre and tone, but its language support is limited |
| Zonos (Zyphra) | A multilingual TTS focused on real-time performance, but falls short of OmniVoice in the number of languages |
Who Is It For
- Video creators and developers who want to produce multilingual narration or dubbing in-house instead of outsourcing each language
- Developers who want to build open-source speech synthesis into voice assistants or game character voices
- People who want to keep speech synthesis entirely in their own environment without sending scripts or audio data outside
- Researchers running comparisons or benchmarks of multilingual TTS
Summary
With its broad language coverage, support for both voice cloning and voice design, and a license that allows commercial use, OmniVoice is a worthwhile option among open-source speech synthesis tools. On the other hand, getting started requires setting up a Python environment, and no support is included. Start by checking the quality for the languages you need with the Hugging Face Spaces demo or the official Colab notebook, and consider installing it locally once you’re satisfied. If you work with someone else’s voice, always obtain their consent first.