github kizuna-ai-lab/sokuji v0.38.0

4 hours ago

Local AI models — per-direction model selection (#436)

  • Each translation direction keeps its own ASR / translation / TTS models. No more flipping the language pair to configure the other side.
  • New "Models" section: one dropdown per stage — Auto shows the model it would actually run, and Browse library… opens the full library scoped to that exact slot and direction.
  • New Storage page for deleting and importing downloaded models.
  • The Models section, warnings, and the session start gate now follow the speaker / participant / both audio mode.
  • The General tab's three overlapping model warnings are now one actionable line, with one-click Switch to Auto.
  • Shorter model names; all new UI translated in every locale.

Fixes

  • Opening the model library for the participant direction froze the whole app (two default-voice writers ping-ponged the Edge TTS voice setting).
  • NeMo streaming ASR models failed to start ("function signature mismatch") — the streaming glue initialized every model as a transducer. Also affected v0.37.x.

Full Changelog: v0.37.4...v0.38.0

Don't miss a new sokuji release

NewReleases is sending notifications on new releases.