Skip to content

MODEL_SETUP.md

Latest commit

 

History

History
40 lines (24 loc) · 1.18 KB

MODEL_SETUP.md

File metadata and controls

40 lines (24 loc) · 1.18 KB

Kanuka Model Setup

Kanuka uses only one model format and one model variant.

1. Required Model File

Select this file in Settings > Model File:

The app validates filename and extension.

2. How to Configure

  1. Open Settings.
  2. Tap Select in the Model File card.
  3. Pick gemma-3n-E4B-it-int4.litertlm from SAF.
  4. Return to the main screen.

The app will auto-preload model/runtime when possible.

3. Runtime Notes

  • Both translation engines use the same .litertlm file:
    • Gemma 3n E4B + SpeechRecognizer
    • Gemma 3n E4B Direct Audio
  • Backend execution is LiteRT-LM on GPU / vision=GPU / audio=CPU.

4. Direct Audio Tuning

In Settings, direct-audio behavior can be tuned:

  • Silence detection sensitivity
  • Skip quickly when no speech is heard (prompt rule)

When the no-speech option is enabled, model prompt instructs early termination behavior if no intelligible speech is detected.

5. Context-Aware Translation

You can enable conversation-context translation on the main screen and configure context turn count in Settings.