What's New
IBM Granite Speech — Local ASR + Direct Speech Translation (WebGPU)
IBM's Granite Speech model (1B parameters) is now available as a local ASR engine with WebGPU acceleration.
- 6 languages for transcription: English, French, German, Spanish, Portuguese, Japanese
- 8 languages for direct speech translation (AST): adds Italian and Chinese
- AST mode: When supported, Granite can translate speech directly without a separate translation model — fewer downloads, lower latency
- Requires WebGPU-capable browser
Participant Mode for Local Inference
Local inference now supports participant audio — translate what other people in the meeting are saying, not just your own voice.
- Automatically selects reverse-direction ASR and translation models
- Falls back to transcription-only mode when translation models aren't available
- Works with system audio capture on supported platforms
Model Preferences Per Language Pair
Sokuji now remembers your model selections for each language pair. When you switch between language pairs, your preferred ASR, translation, and TTS models are automatically restored.
Settings UI Overhaul
- Provider section redesigned with compact model info chips
- Model tags are clickable — jump directly to model selection
- New Help section with support email and GitHub Discussions links
- "API Key" section renamed to "Provider" across all 30 languages
Bug Fixes
- Fixed dismissed tutorials parsing crash in localStorage
- Fixed various model auto-selection edge cases with AST-capable models
Install: Chrome Web Store · Website