What happens to your voice after you approve a capture

When you speak a thought into UTTER IN, three things exist for a short while: the recording, the words in it, and what we understood those words to mean. Only one of them is worth keeping, and it is not the recording.

A recording carries more than you said

A transcript holds your words. A recording holds your voice, the room you were standing in, the television in the background, and anyone else who happened to be talking. None of that is needed to remind you about the dentist on Thursday.

So the rule we built around is short: keep the meaning, not the recording.

What actually happens, in order

1. You record. The audio is uploaded to private storage. It is never public and never served to a browser by link.
2. It is transcribed. A speech-to-text service turns the audio into words. That is the only reason the audio leaves your phone.
3. You see a proposal. A card for each thing you said, showing what we understood. Nothing is saved as a plan yet.
4. You approve. The card tells you, at that moment, that approving keeps what we understood and deletes the original recording. It says so before you press the button, not afterwards.
5. The recording is deleted. What stays is the plan you approved and the text it came from.

And if you never approve?

A recording nobody approves should not linger either. There is an upper limit of 24 hours, and a scheduled job deletes any original recording older than that, whether or not you acted on it. The normal case is much sooner: the recording goes when you approve.

The same number drives both the job and the words on the screen. A promise the interface makes and the system does not keep is worse than no promise at all, so they are not allowed to be two numbers.

Why tell you at the approval step

Because approving is the step you cannot undo. If you wanted to keep a recording, perhaps a voice you care about, you need to know before that moment, not discover it later. A warning after an irreversible action is not a warning.

What we do not do with it

• The recording is not kept "to improve the service".
• What you say is kept out of analytics and crash reports.
• It is not sold, and it is not used to recommend things to you.

The short version

You say something once. We keep what you meant, you check it, and the sound of you saying it is gone. That is the trade we think a voice product should make: less to hold, and less to worry about holding.