OmniVoice
Open-source zero-shot text-to-speech and voice cloning model covering 600+ languages, from the k2-fsa (Kaldi) team, with a paid hosted web app at omnivoice.app
Overview
OmniVoice is an open-source, zero-shot text-to-speech model released in March 2026 by the k2-fsa team, the group behind the Kaldi and k2 speech toolkits. A single model covers 600+ languages (the Hugging Face model card lists 646), including many low-resource languages that commercial voice platforms skip. It supports three modes: voice cloning from a short reference clip, voice design from attributes such as gender, age, pitch, accent and whispering, and an automatic voice. Expressive tags like [laughter] and [sigh] are supported too. The model is Apache 2.0 licensed and runs on NVIDIA GPUs, Apple Silicon and Intel Arc. The omnivoice.app website is a hosted web app built on the model, with pay-as-you-go credits. It is not linked from the official k2-fsa repository, so treat it as a convenience front end rather than the project's official home.
The Verdict
Who Should Use OmniVoice?
Best For
- Creators who need voiceovers in languages ElevenLabs and similar tools don't cover
- Developers who want a free, self-hostable ElevenLabs-style voice cloning model
- Localization projects that need one consistent voice across many languages
- Occasional users who prefer one-time credits over a monthly subscription
Not Ideal For
- Teams that need a polished studio (dubbing editor, audiobook projects, voice library)
- Enterprises that need an SLA, a vendor contract and support from the model's authors
- Non-technical users who want to self-host for free (needs a GPU and Python setup)
What's Great
- Broadest language coverage of any zero-shot TTS model: 600+ languages in one model
- Fully open source under Apache 2.0, commercial use allowed, no per-character fees when self-hosted
- Very fast inference: real-time factor around 0.025, about 40x faster than real time
- Voice cloning from a few seconds of audio, plus voice design from text attributes
- Strong community adoption on GitHub since launch (see badge for count)
- Hosted app uses one-time credits that never expire, so no subscription is needed
Watch Out For
- omnivoice.app is not linked from the official repo; vet it before uploading voices or paying
- Self-hosting needs a capable GPU (or Apple Silicon) and Python know-how
- A research model, not a full product: no dubbing editor, projects or curated voice library
- Quality varies across languages, especially low-resource ones with less training data
- Voice cloning carries consent and misuse risks; there's no built-in voice verification
Pricing
View all features & details
Voice Generation
- Zero-shot voice cloning from a short reference clip
- Cross-lingual cloning (one voice, many languages)
- Voice design from attributes (gender, age, pitch, accent, whisper)
- Auto voice mode
- Non-verbal tags: [laughter], [sigh], [surprise-ah]
Control
- Pronunciation control via pinyin (Chinese) and CMU phonemes (English)
- Speaking speed adjustment (0.5x to 2.0x on omnivoice.app)
- Single-stage, non-autoregressive diffusion language model architecture
Hardware
- NVIDIA GPUs (CUDA)
- Apple Silicon (MPS)
- Intel Arc GPUs (XPU)
Ways to Try It
- Official Hugging Face Space (free demo)
- Official demo page with samples
- omnivoice.app hosted web app (paid credits)
How It Compares
| Feature | OmniVoice | ElevenLabs | PlayHT | Resemble AI |
|---|---|---|---|---|
| Languages | 600+ | 32+ | 142+ | 24+ |
| Open Source | Yes (Apache 2.0) | No | No | Partial |
| Voice Cloning | Zero-shot | Instant & Pro | Yes | Yes |
| Dubbing / Studio | No | Yes | No | No |
| Pricing | Free self-hosted; one-time credits | From $5/mo | From $31/mo | From $25/mo |
| Best For | Rare languages, self-hosting | Quality-first production | Multilingual | Custom voices |