- Fixed a rare bug where switching STT providers quickly could leak the loaded model in memory (up to 2 GB for large Vosk/Whisper models).
- Improved stability of the silence detection stop: the plugin no longer misses stop events under concurrent usage.
- No visible feature changes — this release is a code quality improvement with the bug fixes above.
- Fixed STT provider cache not invalidating when Base URL or model name changes in settings.
- Fixed redundant settings object being passed to
SttProviderFactory(settings are a singleton).
- Added Whisper Local backend — runs a ggml Whisper model on-device via whisper.cpp JNI, no internet required.
- Improved audio resampling quality: switched from linear interpolation to a sinc-based algorithm (TarsosDSP) for better transcription accuracy with Whisper Local.
- Fixed hallucination filter for Whisper Local — added missing Russian subtitle phrases (
субтитры создавал,продолжение следует,dimatorzok). - Refactored offline STT providers: extracted shared lazy-loading logic into
OfflineSttProviderbase class.
- Initial release of SpeakIDE!
- Added two STT backends: OpenAI Whisper (Cloud) and Vosk (Offline).
- Introduced auto-stop on silence detection.
- Added seamless integration with AI Assistant / Junie chat interfaces.
- Added clipboard fallback when no editor is focused.