What happens to the audio

Listening is a request to our server on most plans. Each utterance the voice detector closes is sent up, comes back as text, and no copy of the clip is kept at either end. A tier set to Azure real-time instead sends the audio straight to Azure — our server only hands the app a short-lived token — and it is not kept there either. Either way, the session stores the text.

That is the trade, stated rather than buried: a call needs a network connection, and the audio goes up to become text.