OmniVoice

This first-party Soniqo page documents OmniVoice from the local speech-swift / speech-core implementation. Hugging Face bundles are linked below after the integration notes.

Internal Page First

Landing cards and docs menus now point here first; source model and bundle links remain available from this page.

At a Glance

Model	OmniVoice
Role	Massively multilingual zero-shot voice-cloning TTS
Backend	MLX int8 default bundle; fp16 bundle available
Output	24 kHz mono waveform
Languages	600+ languages
License	Apache-2.0 upstream family
Status	Programmatic speech-swift runtime used by Studio sidecar
Source	k2-fsa OmniVoice
Swift product	`OmniVoiceTTS`
CLI / runtime	Programmatic runtime; not a primary speech speak engine yet

Use

The snippet below mirrors the current speech-swift API or command exposed by the repo.

import OmniVoiceTTS

let model = try await OmniVoiceTTSModel.fromPretrained()
let pcm = try model.generate(
    text: "A new sentence in the reference speaker's voice.",
    referenceAudio: URL(fileURLWithPath: "reference.wav"),
    referenceText: "This is the reference voice.",
    language: "en"
)

Model Links

Implementation Notes

Download repairs incomplete caches by checking the backbone, tokenizer files, and audio_tokenizer model before loading.
Generation iteratively unmasks eight acoustic codebooks with classifier-free guidance.
The runtime combines a bidirectional Qwen3 backbone with a Higgs-audio v2 codec encoder/decoder.
Optional instructions cover restricted style controls such as accent, age, gender, pitch, and whisper.