Trace a spoken request#
Capture audio, produce local text, review the draft, optionally use text refinement and deliberately send the prompt to the selected task. Sending the prompt or an attachment is a separate data transfer from local recognition.
Model preparation and software updates can require network access even when transcription itself is local. Do not use a successful offline transcription test to claim that downloading, refinement or Codex execution is offline.
Retention is a separate question#
“Audio remains local” does not establish a complete local retention schedule or prove that recordings never exist on disk. The reviewed sources do not supply a finalized public audio-retention policy, so this page does not invent one.
Use non-sensitive test phrases while assessing the workflow. If you need an approved handling policy for work material, confirm the local history behavior and any service processing before use. Audio location, text destination and long-term retention should each have evidence matching their claim.