OmniVoice

OmniVoice is a zero-shot text-to-speech model developed by Xiaomi’s speech research team (k2-fsa) and introduced in 2025. It supports 646 languages from a single unified model — approximately 20× more language coverage than ElevenLabs and 5× more than PlayHT — with particular depth in low-resource languages that most commercial and open-source TTS systems fail to support adequately.

Beyond language coverage, OmniVoice introduces two capabilities that distinguish it from other open TTS models: zero-shot voice cloning from a short reference clip, and voice design — generating a new speaker persona from a text description of desired vocal attributes, without requiring any reference audio at all.

Architecture

OmniVoice uses a hybrid diffusion language model architecture that is neither pure autoregressive nor pure diffusion. It combines:

• Diffusion-based synthesis for output quality and speaker fidelity
• LLM-style inference for generation speed and sequence coherence

This combination is what allows OmniVoice to achieve both high naturalness and very low latency simultaneously — a tradeoff that typically favours one at the expense of the other. The model achieves a real-time factor (RTF) of 0.025, meaning it synthesises 40 seconds of audio per second of compute.

Performance

On a 24-language multilingual benchmark:

OmniVoice outperforms ElevenLabs on both accuracy and voice fidelity across the benchmark languages — including non-English languages where ElevenLabs has historically been strongest.

Key Capabilities

• 646-language coverage — including low-resource African, Asian, and Indigenous language families
• Zero-shot voice cloning — clone any speaker from a reference clip of a few seconds
• Voice design — specify a voice using natural language attributes (see below)
• 40× real-time synthesis — RTF 0.025 on standard hardware
• Apache 2.0 license — open for research and commercial deployment

Voice Design: Generating Speakers Without Reference Audio

Most voice cloning systems require a recording of the target speaker. OmniVoice introduces voice design as an alternative: you describe the voice you want using natural language attributes, and the model generates a speaker persona matching that description.

Attributes you can specify include:

• Gender and approximate age
• Pitch register (high, mid, low)
• Speaking rate (slow, natural, fast)
• Accent or regional dialect
• Stylistic character (formal, warm, authoritative, whispered)

This capability is practically significant for product teams that do not have a reference recording — for instance, when creating a branded AI persona, generating character voices for interactive media, or building voice interfaces for new languages where no existing recordings exist.

Enterprise Relevance

OmniVoice is most relevant for:

• Global product localisation — voice output in 646 languages from a single model, eliminating the need for per-language TTS vendors
• Low-resource language markets — serving users in languages that commercial APIs do not support
• Privacy-sensitive deployments — on-premise inference under Apache 2.0, no external API dependency
• Branded voice creation — voice design removes the dependency on recording sessions for persona development
• Regulated industries — healthcare, legal, and government applications requiring data sovereignty

References

• OmniVoice paper – arXiv:2604.00688
• OmniVoice GitHub (k2-fsa)
• OmniVoice demo page
• OmniVoice vs ElevenLabs – HM.AI