v4.1.72026-09-12macOSWindowsLinux
More reliable local dictation and model downloads
Give slower local models time to finish, prevent unreliable Windows key polling from cutting recordings short, and respect model-server retry delays.
- Finish longer local processing. Local transcription can finish beyond one minute while cancellation remains available.
- Steadier model downloads. Rate-limited model downloads wait before retrying and preserve server cooldowns.
Changed
- Keep bounded recording-stage and hotkey timing diagnostics without collecting device names or dictated content.
- Separate cancelled dictations and empty-audio outcomes from failed dictations in anonymous analytics.
- Explain provider rate limiting separately from a general cloud service failure.
Fixed
- Allow local transcription up to ten minutes instead of cancelling at one minute.
- Require observed Windows key-down state before the release watchdog can stop a recording.
- Respect model-server Retry-After delays and avoid retrying permanent HTTP failures.