In order to get a text-only transcription you need to record it somehow. I can't be sure that a backup audio file isn't stored anywhere, so I believe the same applies. It means that in theory, your conversation with someone else may be shared. I don't consent to that, normally. For presentations, talks etc it's a different matter.
If you store captions produced by a party you trust (e.g., Google Meet), it could be accomplished w/o accessing audio at all. If I get it right, your concern is more about what's in the pipeline than what comes out.