Can ChatGPT transcribe audio?
Yes, if the audio is happening now. ChatGPT Record captures and transcribes meetings and voice notes — with speaker separation and a 4-hour cap per session — on Plus, Pro, Business, Enterprise and Edu, and only in the macOS desktop app. Dictation transcribes your own speech on every plan. What is missing is the obvious one: audio files are not in the supported upload types.
Three different features hide behind one question.
ChatGPT Record is the real transcription product. Press Record in any chat and it live-transcribes, with a timer, pause and resume; on finishing it uploads a transcript and generates notes as a canvas you can turn into an email, a plan or code. It handles multiple speakers, labelling unknown ones as Speaker 1 and letting you rename them afterwards. Sessions cap at 4 hours, after which recording stops automatically and notes are generated anyway. It is available on Plus, Pro, Business, Enterprise and Edu — and only in the macOS desktop app. Enterprise and Edu workspaces have it off by default until an owner enables it.
Dictation is the everyday one, on all plans: you speak, it types. OpenAI shipped a new speech-to-text model for it across every tier, reporting word error rate at least 10% lower than the previous production model for the top languages tested, with gains in Japanese, Korean, Chinese, Urdu, Vietnamese, accented English and long-form Spanish, French, Italian and Portuguese, plus better handling of noisy rooms and quiet speech. Those are OpenAI's own evaluation figures, not ours.
Uploading an audio file is the gap. OpenAI's supported upload list covers text files, spreadsheets, presentations and documents — XLSX, XLS, CSV, TSV, DOCX, PPTX, PDF, TXT — and image inputs accept PNG, JPEG and non-animated GIF. No audio container appears in either. So the most intuitive version of this task, dragging in a recording you already have, is the one route that does not exist in ChatGPT.
Two honest caveats, both OpenAI's. Record "works best in English today" with accuracy for other languages varying. And OpenAI puts a consent warning in its own article — check local law and get consent before recording others — which we repeat here rather than leave to the small print, because the legal exposure sits with the person pressing the button.
What happens to the recording.
This distinction is the one to understand before recording anything sensitive, and OpenAI states it clearly.
- Audio recordings are used for transcription only and deleted afterwards. They are not used to train models.
- Transcripts and canvases are different. On Free, Plus and Pro accounts with "Improve the model for everyone" enabled, OpenAI may use them for training — and notes that reviewing them may be necessary when doing so.
- Business, Enterprise and Edu workspace content, including Record transcripts and canvases, is excluded from training by default.
So the privacy control that matters here is the same training toggle described on our chat privacy page. Deleting the conversation removes the canvas and transcript within 30 days.
OpenAI's article carries an explicit warning — use responsibly when recording others, check local laws, and get the right consents — and adds that responsibility for lawful use sits with you.
That is not boilerplate. Recording-consent law varies sharply: some jurisdictions require every participant to agree, others only one. A meeting-notes tool makes it trivially easy to record a call without anyone noticing, which is exactly why the warning exists. Say you are recording.
Record can auto-join scheduled calls, with a constraint worth knowing before you rely on it: calendar auto-join supports Google Meet links only. Google Calendar has to be connected in ChatGPT so Record can sync upcoming meetings, and the event must use a Meet link.
For Zoom, Microsoft Teams, Webex or anything else, you start Record manually in the macOS app. And if the Record button is missing entirely, OpenAI's troubleshooting is to update the macOS app and confirm you are logged into a plan that includes it — plus grant both Microphone and Screen & System Audio Recording permissions in System Settings.
Since ChatGPT will not take an audio file, the practical path is to transcribe elsewhere and bring the text in. OpenAI's own speech-to-text models are available through the API, which is the same family of models behind dictation, and there is a healthy market of transcription tools besides.
Once you have text, ChatGPT is genuinely good at the part that follows — summarising, extracting decisions and actions, rewriting into an email or a plan. That is a text task, and it works on any plan including Free.
Routes that accept a file.
ChatGPT Record is built for live capture on one operating system. These cover the cases it does not. We have not benchmarked transcription accuracy across any of them.
Frequently asked.
Quick follow-ups people search after this question.