What happens to the audio you record
Every dictation goes to the VCode machine to be transcribed. The clip is not saved anywhere and the transcript is never written to a log.
Why you would use it
You are about to dictate something you would not paste into a third-party web form. This is the path the audio takes, and the checks in the code that keep it short.
How to use it
Nothing to do. This describes what already happens. There is no setting that keeps the audio on the phone: the machine transcribes every dictation.
What you see
Nothing extra in the UI. What you can check:
journalctl --user -u v-codeshowstranscribed 48213b clip.webm → 112 chars: sizes only, never the clip and never the words.journalctl --user -u v-code-voiceshowstranscribe audio=7.4s took=0.62s status=200: the same shape, no text.
Limits and known gaps
- Anything that can log in to VCode can record and transcribe as you. Login is full trust; see trust-model-and-security-headers.
- The sidecar has no auth and no TLS. It stays on loopback, and anything running as this user on that machine can post a clip to it.
- If you point
VCODE_STT_URLat someone else's service, none of this applies to what that service does. What VCode does not do is store or log the clip on the way there. - Two removed dictation modes kept speech models in Cache Storage on the phone:
portal-live-engine(Live text) andtransformers-cache(an in-browser Whisper engine). Those were models, not your audio. The next app update removes both.
Related
- install-the-voice-sidecar: the service that does the transcribing
- point-dictation-at-another-service: sending clips somewhere else
- dictate-a-message: the recording this describes
- trust-model-and-security-headers: what logging in is worth
- logs-and-troubleshooting: the journal lines quoted above