OpenAI currently documents audio uploads for paid ChatGPT subscriptions and workspaces, including Enterprise. Free-plan audio uploads are not currently supported, and access can depend on workspace settings, region, client version, and selected model. Audio upload availability
This is for a recording you already have. ChatGPT Record captures a live session in its supported app, while dictation turns a spoken prompt into editable input.
Attach the recording and state the transcription rules
Use a recording you have permission to process in the chosen account. Keep the original file and give the upload an identifiable name. If you supply only an excerpt, record its start point in the original so you can find a disputed passage later.
OpenAI lists formats including WAV, MP3, FLAC, AAC, and M4A. Audio-only MP4 and WebM are supported, but files identified as video are not supported by audio uploads. A file must contain a decodable audio stream. Renaming a video extension does not change its contents. Supported audio formats
Tell ChatGPT whether you want literal speech or a lightly cleaned reading copy. For a quote you may use in a report, start with the literal version. Removing a repetition can improve readability; deleting “not” or an unfinished qualification changes the claim.
Transcribe the attached [FILE NAME] in its spoken language.
Purpose: [REVIEW / INTERVIEW QUOTES / OTHER].
Preserve the spoken wording, including qualifications, negations,
and corrections. Do not summarize, translate, or polish the speech.
Mark unclear words [inaudible] rather than guessing.
Separate speaker turns where you can hear a change. Use Speaker 1,
Speaker 2, etc., unless identities are explicitly established.
Do not infer names from voice characteristics. Flag uncertain turns.
Include approximate elapsed timestamps at useful review points if
you can establish them. Never invent timestamps. State any portion
you could not process and the last reliably transcribed point.
Return the transcript, then a separate list of uncertain passages.
This is [THE FULL RECORDING / AN EXCERPT BEGINNING AT ORIGINAL TIME].
Keep excerpt elapsed times separate from original recording times.
These are requested output rules, not guarantees of accurate timestamps or speaker labels. OpenAI warns that transcripts can contain errors and speaker identification can be unreliable. Review against the audio.
Replay the words that change the meaning
This fictional passage is an authored practice example, not recorded speech or a ChatGPT output:
Speaker 1: We received fifteen cases, not fifty. The replacement is not approved yet.
A transcript reading “We received fifty cases. The replacement is approved” would change both the quantity and the status. Check numbers, identifiers, names, negations, and permission or approval language before correcting punctuation.
Use the playback position to locate the passage yourself. If a timestamp points nearby, listen before and after it. If the voices overlap and the speaker remains unclear, retain a neutral speaker label or ask someone with direct knowledge to identify the turn. A familiar topic does not establish who said it.
Keep corrections separate from editorial cleanup
After checking the transcript, make a reading copy if you need one. Preserve the original transcription and tell ChatGPT exactly what cleanup is allowed:
Make a separate reading copy of this checked transcript.
Remove filler words and repeated starts only when meaning is intact.
Keep quantities, negative statements, uncertainty, corrections,
and speaker labels. Do not resolve remaining [inaudible] markers.
List any edit that could affect a quotation's meaning separately.
Checked transcript: [PASTE]
For the fictional passage, “fifteen, not fifty” is a correction carrying information. Keep it in a literal quote even if it looks repetitive. Check the recording again before presenting an edited sentence as someone's exact words.
Handle an incomplete upload without inventing the rest
OpenAI lists a 512 MB audio-file maximum. Longer recordings may be processed in smaller sections when Data Analysis is available, and very long recordings may time out. A file below the size cap is not a guarantee that every minute was processed. Audio processing limits
Compare the transcript's ending with the recording's ending. If material is missing, provide a permitted shorter section and keep a mapping to the original timeline. If an upload fails, check the actual format and audio stream, account access, and any displayed error before retrying.
Save the checked transcript with the recording name, excerpt mapping, and unresolved passages. Anyone using a quotation should be able to return to the corresponding audio, including the qualification that came immediately before or after it.