How Whisper dictation handles audio
SHORT ANSWER
With Whisper enabled, the paired computer’s daemon reassembles the recording and runs local whisper.cpp. It does not log audio or transcript content. Temporary-file deletion is best-effort, so the implementation does not promise that cleanup can never fail.
Where transcription runs
- The app base64-encodes the completed recording and sends indexed
AudioChunkframes to the paired computer. - The daemon rejoins and decodes those chunks, then invokes the computer’s local
whisper.cpppipeline. - The transcription result returns to the device that sent the recording; it is not injected into a separate agent process.
Logs and temporary files
- The implementation does not write audio bytes or transcript text to daemon logs.
- The normal CLI path logs the exit code and duration. If transcription throws, the error path logs the exception type and up to 120 characters of its message.
- Audio and intermediate output live under a temporary directory. A
finallyblock attempts recursive deletion after the run.