Transcribe recorded audio

The technical side of dictation: what the service accepts, what it costs, and what it returns.

Limits

Recording stops automatically at the duration limit, and the audio captured up to that point is transcribed rather than discarded.

The size limit is generous for dictation. POM AI records at a low bit rate suited to speech, so three minutes is well under a megabyte in practice — the ceiling exists to reject anything abnormal, not to constrain normal use.

Cost, and why short recordings are poor value

Cost is calculated per started minute. Any recording up to sixty seconds costs six credits; up to two minutes, twelve; the full three minutes, eighteen.

That rounding is the one thing worth planning around. Five ten-second recordings cost thirty credits and contain less than one fifty-second recording costing six.

Dictation is also the only POM AI operation that charges without an estimate or a confirmation dialog. It transcribes as soon as you stop recording. The exposure per recording is bounded at eighteen credits, but it is not something you approve first.

See What POM AI credits are.

Audio formats

The recording format is chosen by your browser from the candidates POM AI offers, preferring Opus in a WebM container and falling back through Ogg, MP4 and MPEG as support allows.

The service accepts WebM, Ogg, Opus, MP4, MPEG, WAV, M4A, AAC and 3GPP audio, along with WebM and MP4 video containers, since browsers sometimes label an audio-only recording with a video container type.

You never choose the format. It is negotiated between your browser and the recorder, which is why dictation works on a phone and a desktop without configuration.

What it accepts and what it does not

Dictation transcribes a recording made in the browser at that moment. There is no upload field, and no way to submit an existing file: no meeting recordings, no podcasts, no voicemail, no video.

If you need those transcribed, this is not the tool.

Validation before the recording is sent

Your site checks the recording before forwarding it, so an unusable one fails immediately rather than after an upload:

• the duration is within the limit;
• the recording is not empty and not over the size limit;
• the audio is genuinely one of the accepted types, determined by inspecting the file's real content rather than trusting the label the browser attached.

Failing these produces a message in the status line and costs nothing.

What the service returns

On success, the response carries the transcript, the language detected when the provider supplies one, the duration accepted, and the credits consumed. The transcript is then appended to your prompt field.

The usage log on the account records the operation as transcribe_audio against the domain that requested it, so dictation spending is visible alongside everything else. See Review your credit history.

Requirements

Transcription is an account service, so it needs the same things every other POM AI operation needs:

• a valid license with this domain authorized;
• available credits, and a domain spending limit that has not been reached;
• the ability for your server to reach the account host.

A dictation failure that coincides with other POM AI tools failing is a licensing, credit or connectivity problem rather than an audio one. See Troubleshoot license and domain errors.

Extending the limits

The duration, size and accepted-format limits are filterable, so a developer can adjust them for a specific installation. Raising them beyond what the account service accepts does not help — the account applies its own limits and will reject a recording your site allowed through.

Treat the documented values as the real ceiling.

When transcription is refused

See Troubleshoot dictation.