Last updated: 11/18/2026Last verified: 2026-11-15

OmniVoice

OmniVoice Overview: Cloning, Voice Design and 646 Languages

5
0

OmniVoice is an Apache-2.0 zero-shot text-to-speech model from the k2-fsa speech group, built on a diffusion language model architecture: clone a voice from a short reference clip or design one from a description, in 646 languages and dialects according to the project’s language list, with public weights and a Gradio demo on Hugging Face. It runs in Python on CUDA, Apple Silicon and Intel Arc hardware.

Free

What is OmniVoice?

OmniVoice is the k2-fsa speech group's open-source zero-shot text-to-speech model, built on a diffusion language-model-style architecture with a Qwen3-0.6B backbone. Give it a short reference clip and it clones the voice; describe a voice in words and it designs one. Its language list documents 646 languages and dialects over 581,000 hours of training data, the weights are public and ungated at about 3.27 GB in total, and it runs in Python on CUDA, Apple Silicon MPS and Intel XPU. There is no official product site: the project lives at github.com/k2-fsa/OmniVoice, with a Hugging Face Space and a Colab notebook for trying it. This listing read the README, docs/languages.md, LICENSE, the tokenizer licence, pyproject.toml, the Hugging Face model page and the GitHub API on 2026-11-15, but did not install the dependencies or synthesize speech.

Key Features

  • Zero-shot voice cloning from a short reference clip, without training a speaker model.
  • Voice design from a description, covering gender, age, pitch, dialect or accent and whisper; the README notes that mode was trained on Chinese and English data only.
  • 646 languages and dialects per docs/languages.md, drawn from 581,000 hours of training data, while the README rounds the figure to 'over 600'.
  • Public, ungated weights: a 2,450.3 MB model.safetensors plus an 805.7 MB audio tokenizer, about 3.27 GB in all, on a Qwen3-0.6B base.
  • Runs on CUDA, Apple Silicon MPS and Intel Arc XPU; the project reports testing a 4 GB Arc A310 and a 16 GB Arc Pro B50.
  • Performance figures from the project: real-time factor 0.0899 on a single H100 in fp16, and 0.0115 with FlashInfer at batch size 8.
  • Entry points include a PyPI package, three CLI tools, a Gradio demo, a Hugging Face Space and a Colab notebook; the paper is arXiv:2604.00688.

Best For

Developers building multilingual speech synthesis into something they can host themselves.Researchers experimenting with zero-shot cloning and description-driven voice design.Teams that want speech generation without per-character billing.Prototypes that need wide language coverage more than studio polish.Anyone with an NVIDIA, Apple Silicon or Intel Arc GPU who would rather run weights locally.

Pricing

OmniVoice is free to use: the repository is Apache-2.0 and the weights download without a gate or an account, so self-hosting carries no per-character fee. You still pay for your own GPU or cloud compute, and the Hugging Face demo and any third-party host follow their own terms. One licence detail matters commercially: inference loads the Higgs Audio V2 acoustic tokenizer, which ships under the Boson Higgs Audio 2 Community License rather than Apache-2.0, and that licence requires an expanded licence from Boson AI once a product passes 100,000 annual active users.

Pros & Cons

Pros

  • Both zero-shot cloning and description-driven voice design in one model.
  • Public, ungated weights that run on consumer and workstation GPUs alike.
  • A documented 646-language list rather than a marketing claim.
  • Several entry points: PyPI, CLI, Gradio demo, Hugging Face Space and Colab.
  • Published inference speeds, so throughput can be estimated before deploying.
  • No per-character fee when self-hosted.

Cons

  • Two licences, not one: the repository is Apache-2.0, but the required audio tokenizer is under the Boson Higgs Audio 2 Community License, whose additional commercial terms apply above 100,000 annual active users, and the Hugging Face model card states no licence at all.
  • Voice design was trained on Chinese and English only, and cross-lingual cloning carries the reference clip's accent; cloning is the more stable mode.
  • The README publishes no VRAM table, so sizing a deployment is guesswork.
  • On macOS the pynini dependency ships no wheel and has to come from conda.
  • It is a model and a Python project, not a hosted product: no dashboard, SLA or support contract.

Alternatives

ElevenLabs

A hosted commercial platform with managed latency and support, instead of weights you run yourself.

Fish Audio

A hosted studio and pay-as-you-go API for text-to-speech and cloning, without local GPU setup.

Soniox TTS v2

A hosted low-latency streaming TTS API with emotion tags, priced per generated hour.

Seed Audio 1.0

ByteDance's hosted model that generates dialogue, effects and ambience in one scene rather than speech alone.

FAQ

Details

Platform

macOSWindowsLinuxAPI

Features

  • Zero-shot voice cloning from a short reference clip plus voice design from a text description, on a diffusion language model-style architecture, with no per-character fee when you self-host
  • Multilingual coverage: the README says “over 600 languages” and links to a language list whose first line reads “OmniVoice supports 646 languages with a total of 581k hours of training data”
  • Public, ungated weights on Hugging Face (model.safetensors 2450.3 MB plus an 805.7 MB audio tokenizer, built on a Qwen3-0.6B backbone), a Gradio Space, a Colab notebook and a PyPI package, with a paper at arXiv:2604.00688
  • Runs on CUDA, Apple Silicon MPS and Intel XPU; the project’s own benchmarks quote real-time factors of 0.0899 on a single H100 in fp16 and 0.0115 with FlashInfer at batch 8

Languages

enzhjakofrdeesruarhi

Known limitations

  • There is no official website: omnivoice.ai is an unrelated business VoIP service, omnivoice.com is for sale, omnivoice.dev is a different WhatsApp voice project and omnivoice.io is parked, so the repository is the only official home and no hosted paid API exists
  • The licence is not plain Apache-2.0 in practice: inference requires the bundled Higgs Audio V2 acoustic tokenizer, whose BOSON HIGGS AUDIO 2 COMMUNITY LICENSE adds commercial terms — more than 100,000 annual active users needs an expanded licence from Boson AI — and the Hugging Face model card declares no licence at all
  • Voice design is trained on Chinese and English data only (the README says it “can generalize to other languages, but may produce unstable results for some low-resource languages”), cross-lingual cloning carries the reference clip’s accent, and the model is most stable on the cloning task it was primarily trained on
  • The README publishes no VRAM table — the only hardware datapoint is an Intel Arc A310 with 4 GB — and on macOS the pynini dependency has no wheel and must come from conda. The 646-language count and the speed figures are the project’s own; the repository held 85 open issues with its last tagged release 0.2.1 on 2026-07-16, and this listing did not install or run the model

Rate This Tool

Related Tools