v3.0.2

deepseek-ai/DeepSeek-OCR-2v3.0.2Aug 25, 2026by gabrieldemarmiesse

Upgrading?

AI Summary

A maintenance release integrating the Czech model and improving training efficiency for the `lsd_scratch` component.

Key Highlights

  • Added Czech model to the README
  • Optimized `lsd_scratch` with cosine schedule and batch size 16x4
  • Default voice changed to 'alba' with voice cloning support

New Features

  • Czech model support
  • Optimized lsd_scratch training parameters

Full Release Notes

## What's Changed
* Add the Czech model to the README and plug the community models section by @vvolhejn in https://github.com/kyutai-labs/pocket-tts/pull/246
* lsd_scratch: 5s voice prompts, cosine schedule, batch 16 x 4 by @manu-lm in https://github.com/kyutai-labs/pocket-tts/pull/247
* Now --voice defaults to alba with voice cloning on custom config by @gabrieldemarmiesse in https://github.com/kyutai-labs/pocket-tts/pull/248


**Full Changelog**: https://github.com/kyutai-labs/pocket-tts/compare/v3.0.1...v3.0.2

Related Repositories