- Improved translation speed for repeated lines by serving them directly from the instant cache.
- Fixed an issue where short phrases such as "Okay" and "Thank you" could be incorrectly detected as music or background noise and skipped.
- Enhanced the noise filtering system to always preserve genuine speech while removing only recognizer loop artifacts ("phantom" subtitles).
- Fixed translation text corruption affecting measurements (50 km/h), fractions (3/4), parenthetical expressions (e.g. "(sighs)"), and legitimate repeated words.
- Fixed an issue where rapidly switching AI models could freeze the loading screen or unload the wrong model.
- Improved recovery when loading large models, preventing empty translation results after initialization delays.
- AI engine processes are now automatically terminate after an unexpected application crash, preventing unnecessary GPU (VRAM) usage.
- The application now manages only AI processes that it started, leaving external local AI servers (such as llama-server) untouched.
- Improved settings storage to prevent corruption or resets after crashes or power loss.
- Fixed native memory leaks in the speech recognition engine during extended sessions.
- Fixed an issue where live captions could become unresponsive after losing connection to the translation engine.