Request: Lightweight / Distilled OmniVoice Variant for On-Device Use

[ajay01994]

Hi team,

Thank you for the excellent work on OmniVoice — the support for 600+ languages and zero-shot voice cloning is truly impressive .

However, the current model is relatively heavy for on-device usage (e.g., ~4GB VRAM on a laptop), which limits real-time and edge deployment scenarios.

I wanted to ask if there are any plans to release a lightweight or distilled version of OmniVoice (e.g., ~100M–300M parameters), similar in spirit to ZipVoice, that:

Supports a smaller subset of languages (e.g., 5–10)
Retains voice cloning capability
Is optimized for edge / real-time inference

Such a model would be highly valuable for deployment with frameworks like Sherpa-ONNX / edge pipelines.

Looking forward to your thoughts.

Thanks!

[zhu-han]

Hi, we do not plan to develop a smaller version of OmniVoice, at least in the short term. However, we will develop an upgraded version of ZipVoice for edge devices in the future.

[ajay01994]

Hi, thanks for the response!

Currently, ZipVoice seems to support only Chinese and English. I wanted to ask if the upcoming upgraded version will be multilingual (closer to OmniVoice), or will it remain limited to a few languages?

Also, do you have any tentative timeline for the release?

Looking forward to it — this would be very valuable for on-device multilingual applications.

[zhu-han]

I'm aiming for a multilingual version. I can't promise a timeline as there are a lot of uncertainties during model development. But I'll try my best to get it done sooner. Thanks for your attention to our work.