Inicio - Documentación - POM AI - 07 Dictation and transcription - Understand audio data handling

Understand audio data handling

Dictation sends a recording of your voice off your server. This page describes what is captured, where it goes and what remains.

What is captured

Audio is captured only while a recording is active — between starting and stopping. The animated level indicator beside the button is the visible signal that capture is happening.

The browser is not listening at other times. There is no wake word, no background listening and no capture while the page merely sits open.

Recording requires an explicit microphone permission you granted to the site in your browser, which you can revoke at any time. See Allow microphone access.

Where the recording goes

  1. The browser holds the recording in memory while you record.
  2. On stop, it is uploaded to your own WordPress site.
  3. Your site validates it and forwards it to the POM AI account service, authenticated with your license and domain.
  4. The service returns a transcript.

Two hops, both server-to-server after the first. The recording is not sent anywhere else by POM AI.

What is stored

On your site: nothing. The recording is handled during the request and is not written into the media library, the uploads directory or the database. There is no attachment, no post and no log entry containing audio.

In the transcript: the text you dictated. Once inserted it is ordinary text in a prompt field. If you then generate, that text is sent as part of the prompt like any typed text.

On the account: the usage record. The operation is logged as transcribe_audio with the domain, the credits used and the timestamp. That record covers the fact of the transcription, not its content.

The transcript is not stored by POM AI outside the prompt field you are looking at. If you navigate away without generating, the text goes with the page.

What accompanies the request

The same identification every POM AI service request carries: the license key and the current domain, plus replay-protection values. The recording is sent encoded inside the signed request body so it cannot be altered in transit without invalidating the signature.

Nothing about the visitor to your public site is involved. Dictation happens in wp-admin, initiated by a logged-in editor or administrator.

Never dictate sensitive information

Everything you say becomes prompt text, and prompt text leaves your infrastructure and cannot be recalled. The rules are exactly those for typed prompts:

  • no passwords, API keys, license keys or tokens;
  • no bank or card details;
  • no identity document numbers;
  • no customer records or health information;
  • no unpublished commercial information.

Speaking is more casual than typing, which is precisely the risk. It is easier to mention a customer's name aloud than to type it deliberately. See Keep sensitive information out of prompts.

Other people's voices

Record only your own voice. Recording someone else — a colleague on a call, a customer in the room — captures their voice and their words and sends both to a third-party service.

If a recording captures someone else incidentally, treat it as you would any other unintended personal data: do not generate from it, clear the prompt field, and follow your own procedure if the content warrants it.

Where you dictate matters

Dictation captures whatever the microphone hears, including background conversation. In an open office or a shared space, other people's discussions can end up in the transcript and therefore in the prompt.

Read the transcript before generating — partly for accuracy, partly for this.

Your obligations

POM AI processes what you choose to send it. Deciding what may lawfully be recorded and transmitted, informing anyone whose voice or data is involved, and recording that decision remain yours as the site operator.

If your organisation has a policy on voice recording, dictation falls under it.

If something was said by mistake

You cannot unsend a recording. What you can do:

  1. Clear the prompt field so the transcript is not sent again as part of a generation.
  2. Do not generate from a prompt containing the mistake.
  3. Rotate any credential that was spoken aloud. A license key is rotated from My Account → Licenses & Downloads; see Rotate a license key safely.
  4. Follow your incident procedure if personal data was involved, and tell whoever is responsible for data protection.

Turning dictation off

There is no separate switch, because dictation is part of the tools that have prompt fields. To prevent it on a machine, revoke the site's microphone permission in the browser. To prevent it for a person, remove their access to the tools.

For the wider picture of what POM AI transmits, see Understand data sent for AI processing.