Can ChatGPT Transcribe Audio and Video? (Honest 2026 Answer)

The 3 methods, ranked by usefulness

"ChatGPT transcription" means three different things depending on which product surface you're using. Ranked by how often each one is the right call:

Capability by ChatGPT tier

What each ChatGPT plan can and cannot do for transcription. Verified against OpenAI's ChatGPT Record help article, chatgpt.com/pricing, and MacRumors' 2025-07-17 Plus rollout confirmation.

Common misunderstanding: "ChatGPT can transcribe" conflates the Record feature (macOS desktop live capture) with the GPT-4o file upload (existing audio files, undocumented but functional) with the OpenAI API (developer endpoint, separate product). Different limits, different platforms, different results.

ChatGPT Record mode — deep dive

ChatGPT Record is OpenAI's built-in live audio capture, rolled out to Plus users globally on July 17, 2025. It captures meetings, brainstorms, lectures, and voice notes in real time and produces a transcript plus AI-generated notes.

Availability (verified 2026-09-09)

• macOS desktop app only. Windows support has no announced roadmap. Web ChatGPT does not have Record.
• Plans: Plus, Pro, Business, Enterprise, Edu. Not available on Free.
• Enterprise admins can disable Record per workspace.

Session limits

• 4-hour cap (240 min) per session. Sessions exceeding this stop automatically and generate notes uploaded as a private canvas.
• No hard file-size limit for the resulting transcript (it's stored in the chat).
• Language support: English at launch, expanding gradually.

Output

• Transcript — single-track prose, no speaker labels.
• Summary — auto-generated overview.
• Time-stamped citations — clickable references from the summary back into the transcript. This is the differentiator vs plain transcription tools.
• Suggested follow-ups — action items and next-step prompts.
• All output appears in a canvas you can edit or export.

Privacy

• Audio recordings are deleted after transcription. Only the transcript persists (in your chat history until you delete it).
• OpenAI does not train models on Record audio. Stated directly in the help article: "We do not use these audio recordings to train our models."
• The transcript is a different question from the audio. Transcripts and canvases can be used for training on Free, Plus and Pro where "Improve the model for everyone" is switched on. The audio is deleted either way — the text it produced is what the setting governs.
• Business, Enterprise and Edu are excluded from training by default. Record is also disabled by default on Enterprise and Edu workspaces until an admin turns it on.

Steps

1. Open the ChatGPT macOS desktop app.
2. Start a new chat. Click the Record icon in the input bar.
3. Grant microphone permission. Recording starts. Audio is captured on-device and streamed to OpenAI for real-time transcription.
4. Speak, or let the meeting audio play through your Mac speakers (Record captures system audio too).
5. Click stop. The canvas opens with the transcript, a summary, and suggested follow-ups.
6. Ask ChatGPT to reformat, extract action items, translate, or restructure — all in the same chat.

ChatGPT dictation — the third method, and the usual mix-up

The microphone button in the ChatGPT app is dictation, not transcription. It is available on every tier, Free included, on mobile and desktop. You speak, and your words become the text of the message you are about to send. That is speech-to-input.

It cannot transcribe a file. There is nowhere to attach one — the microphone only listens to your microphone, live, for as long as you hold the conversation. People on the Free tier often find this button, assume they have found audio transcription, and then discover there is no way to point it at the recording sitting on their desk. There isn't one: Free has dictation and nothing else.

It is genuinely good at what it does. Dictating a long prompt is faster than typing it, and the punctuation inference is solid. Just don't plan a transcription workflow around it — for an existing file you need Record (macOS, paid), the GPT-4o upload (paid), the API, or something else entirely.

Transcribing existing audio files (GPT-4o upload)

Since GPT-4o launched in 2024, ChatGPT Plus / Pro / Business accept audio file uploads directly in chat. It processes the audio through Whisper server-side and returns a transcript in the chat. Important caveat: as of 2026-09-09, OpenAI's official "supported file types" documentation lists only "all common file extensions for text files, spreadsheets, presentations, and documents". Audio appears nowhere in it. Upload works in practice but is not formally documented — meaning behaviour could change without notice, and there is no published limit to appeal to when it does. The 25 MB figure below comes from OpenAI's API documentation, which is explicit; the chat window's behaviour with audio is an observation, not a documented guarantee.

What works

• Formats: MP3, WAV, M4A, WebM. MP3/M4A most consistent.
• Size limit: 25 MB per file (matches the OpenAI API limit).
• Duration: ~10-15 minutes of typical-bitrate audio fits under 25 MB.
• Output: prose transcript in chat, no timestamps, no speaker labels.
• Post-processing: same chat can then summarize, translate, extract action items.

What doesn't work well

• Files over 25 MB — hard fail.
• Files under 25 MB but over ~10 minutes — often returns incomplete transcripts without warning.
• Multi-speaker audio — no diarization.
• Accented English, overlapping speech, technical vocabulary — accuracy drops below 80%.
• Non-English audio — supported by Whisper but Record UI + processing prioritizes English.

How to split larger files

For files over 25 MB or longer than 10 minutes, split with FFmpeg into ~10-min chunks:

ffmpeg -i input.mp3 -f segment -segment_time 600 -c copy chunk%03d.mp3

Upload each chunk separately to ChatGPT, then concatenate the transcripts. Or use the OpenAI Whisper API which handles the same 25 MB limit per request but is designed for chunking workflows. For hands-off single-upload processing of long files, a hosted transcription service handles this natively.

Transcribing video with ChatGPT

No ChatGPT tier plays video files directly. Not Free, not Plus, not Pro. Uploading an MP4 to a chat either fails or extracts nothing useful. The OpenAI API accepts MP4 but transcribes only the audio track — the visual content is ignored. Every "ChatGPT transcribes video" workflow is really a two-step: extract audio, then transcribe audio.

Standard workflow

1. Extract audio from the video:
   • FFmpeg (free CLI): ffmpeg -i video.mp4 -vn -c:a copy audio.m4a
   • yt-dlp for YouTube: yt-dlp -x --audio-format mp3 [URL]
   • QuickTime on Mac: File → Export As → Audio Only
   • Online: audio-extractor.net, Kapwing extract-audio
2. Verify the audio file is under 25 MB and ~10 min.
3. Upload to ChatGPT Plus/Pro/Business, or use Record playback capture, or send to the OpenAI API.
4. Ask ChatGPT to reformat, summarize, extract action items.

Shortcut — skip the two-step

Dedicated transcription services read the video directly: VexaScribe accepts MP4/MOV/MKV/WebM/AVI up to 200 MB with no account (5 GB once you have a free one), extracts the audio internally, and returns the transcript with speaker labels — one upload, one download, no CLI.

The realistic workflow — post-hoc, not live

Every ChatGPT-transcription review that actually tested the product (tldv, meetgeek, jotme, umevo) lands on the same conclusion: the realistic workflow for a meeting is after the fact. Record the call somewhere else, export the file, upload it, and hope it comes back usable.

Record mode helps in a narrow band: solo brainstorms captured on macOS, single-speaker interviews under 4 hours, voice notes. It falls apart for the workflows people actually want to use it for.

Where Record actually shines

• Solo brainstorm or voice memo on a Mac (single speaker, English, under 4h)
• 1-on-1 interview where you don't need speaker labels (mark them by hand later)
• Short lectures or study capture where AI-generated notes are the primary output
• Ideation sessions where the summary matters more than the raw transcript

Where Record falls apart

• Multi-speaker meetings — no diarization means everyone becomes one voice
• Long recordings over 4h — session auto-stops mid-flow
• Non-English audio — support is expanding but launch was English-only
• Windows / mobile — no support as of 2026-09-09
• Batch processing — no CLI, no API, no queue
• Zoom/Teams/Meet calls — Record captures system audio locally, but purpose-built meeting bots (Otter, Fireflies, Fathom) integrate directly with the calendar and produce speaker-labeled notes

Prompts that work with ChatGPT + a transcript

ChatGPT's real strength is post-processing existing transcripts — cleanup, formatting, summarization, translation. It is not the transcription engine; Whisper (or whatever ASR you used) is. Five prompt patterns that reliably produce useful output:

You will receive an unpunctuated transcript. Add standard sentence punctuation, capitalize proper nouns and sentence starts, and break the text into paragraphs where the speaker changes topic. Do NOT change any words. Do NOT add facts. Return only the cleaned text.

This is a two-person interview transcript. Label each turn with Speaker 1: or Speaker 2:. If you cannot tell who is speaking, mark the turn [SPEAKER ?]. Do not invent dialogue.

Below is a transcript that references these terms: [Vercel, Next.js, Turbopack, App Router]. Correct any misspellings of these terms to match this exact casing. Leave all other text unchanged.

In this transcript, find any sentence that looks garbled, contains repeated tokens, or has numbers/proper nouns that seem out of context. Wrap each in [FLAG: reason]. Return the full transcript with flags inline.

From this meeting transcript, list every action item as: '[Owner if named]: [action], because [1-sentence quote from transcript]'. If owner is unclear, mark [UNASSIGNED]. Do not invent action items — only extract what is explicitly discussed.

Prompts that fail (and what to do instead)

Common requests that don't work — and the actual working path for each.

File-size, duration limits + OpenAI API pricing

Limits by interface

OpenAI API pricing (transcription models, verified 2026-09-09)

Per-minute billing, no minimum, no publicly-listed volume discount. All of these cap requests at 25 MB per file — the same ceiling as the chat upload, and the reason long-file workflows need chunking whichever route you take. Test on your own audio before standardising: model behaviour differs by length and content type, and the cheapest model is not always the one that needs least cleanup afterwards. See how accurate is Whisper for third-party comparison data.

Accuracy — honest limits

This page used to carry a table of accuracy percentages. It no longer does, and the reason is worth stating plainly: we could not stand behind the numbers. They came from third-party tests that did not publish a methodology we could inspect, on model versions that OpenAI has since replaced. A figure like "80-86%" looks precise and rigorous; in practice it was a number copied between articles, ours included, without anyone re-running the test.

What can be said honestly is this. Word error rates on clean, single-speaker audio have fallen far enough across every serious engine that accuracy is rarely the thing that decides the outcome any more. What decides it is structure: whether you get speaker labels, whether you get timestamps, whether the file gets in at all. Those are checkable facts rather than contested measurements, so those are what this table compares.

Where ChatGPT does still lose ground is the hard stuff every engine finds hard: heavy accents, crosstalk, technical vocabulary, poor microphones. The difference is that a dedicated tool hands you speaker-separated text you can correct, while ChatGPT hands you a single block of prose in which you cannot tell who said what. The only test that settles it is your own audio through both.

Privacy — who keeps your audio, and for how long

Recordings are among the most sensitive files people upload anywhere: client calls, medical consultations, interviews with named sources, family conversations. "Does this get used to train a model" is a fair question to ask of any transcription tool, this one included. Here is what each side documents.

OpenAI's documented position

• Record mode audio is deleted after transcription. Only the transcript stays, in your chat history until you delete it — per the ChatGPT Record help article.
• OpenAI does not train on Record recordings. Business, Enterprise, and Edu carry contractual non-training defaults. Plus and Pro follow standard ChatGPT privacy settings, where training on your content is a setting you can switch off.
• The API is separate from ChatGPT. API inputs are not used for training by default — a different policy from the consumer chat interface, and a common source of confusion when people cite one to justify the other.
• Files you upload to a chat are not Record audio. An audio file attached to a conversation is governed by your normal ChatGPT data controls, not by the Record deletion behaviour above.

Ours

• We do not train models on your uploads. Not on the audio, not on the transcripts.
• Your transcripts stay while your account does. They are yours to keep, and yours to delete individually at any time. Delete the account and we remove your personal data within 30 days — the full terms are on our privacy page, which is the authoritative version of all of this.
• The free preview needs no account. There is nothing to sign in to and nothing tied to your identity, because we never asked for it.

If the recording is genuinely sensitive — privileged, medical, or covered by an NDA — the safest answer is not any hosted tool, ours included. It is running Whisper locally, where the audio never leaves your machine. That is slower and needs a capable computer, and for a great many recordings it is more caution than the situation calls for. But when it matters, it is the option with the fewest parties involved.

When to skip ChatGPT for transcription

ChatGPT is a good post-processor and a fine live-recording tool for macOS Plus/Pro/Business users. It is not the right first choice when any of the following applies:

• You have existing audio or video files over 25 MB. Use a dedicated service or the OpenAI API with chunking.
• You need speaker labels. No ChatGPT interface produces them — Record and file upload both return single-track prose. OpenAI’s API has one diarizing model, gpt-4o-transcribe-diarize, but it was deprecated on 2026-08-26 and is removed in February 2027, which makes it a poor thing to build on now. Dedicated services (VexaScribe, Otter, Rev, AssemblyAI, Deepgram) do it as standard.
• Your audio is longer than ~2 hours. API caps requests at 25 MB; Record caps at 4h. Neither handles 8h all-hands or day-long court testimony gracefully.
• You need SRT / VTT / word-level timestamps. ChatGPT output is prose; dedicated tools produce subtitle formats.
• You're on ChatGPT Free. Record isn't available; you're restricted to pasting existing transcripts.
• You're on Windows. Record has no announced Windows roadmap.
• You're processing high volumes. Per-minute API pricing is cheap for individuals, expensive for batch operations; self-hosted Whisper or a hosted subscription is more predictable.
• Meeting transcription with participants. Purpose-built meeting bots (Otter, Fireflies, Fathom, tldv) integrate with your calendar, join Zoom/Teams/Meet as a participant, and produce speaker-labeled notes.

What a dedicated tool adds

Not accuracy — that argument is mostly over, and the accuracy section above explains why we stopped making it. What a purpose-built tool adds is structure and the things you do with a transcript after you have it. In our case:

• Automatic speaker labels — multi-speaker, renameable in the editor — and timestamps on every transcript and every export.
• Exports: TXT, SRT, VTT, DOCX, CSV, JSON. Word-level timestamps come in the JSON export.
• AI summary in meeting, sales, lecture and podcast formats — the one thing every competitor on this topic also does.
• AI chat — ask questions, answers cite the timestamp. This is the ChatGPT habit you already have, pointed at a transcript that knows who spoke and when.
• Translate the transcript in place — free translations included; premium keeps terminology consistent across long recordings.
• Bulk upload and bulk export, an editor with audio-synced playback, and 99 languages detected automatically.

No per-minute billing, no chunking, no prompt engineering. Full detail on the audio transcription page.

Verified sources

Every capability, limit, and price on this page was re-checked against these sources on 2026-09-09:

• OpenAI Help Center — ChatGPT Record (feature, tier availability, privacy)
• chatgpt.com/pricing (Free / Plus / Pro / Business / Enterprise / Edu plans)
• MacRumors — ChatGPT Plus Gets Record Mode on Mac (July 17, 2025 global Plus rollout)
• AIbase — ChatGPT Voice Recording Mode Now Fully Available (macOS-only scoping)
• OpenAI API pricing (per-minute rates for every transcription model)
• OpenAI API deprecations (the 2026-08-26 transcription-model deprecations and their 2027-02-26 shutdown)
• OpenAI speech-to-text guide (25 MB request limit, supported formats, the gpt-transcribe recommendation)
• Accuracy percentages were removed from this page on 2026-09-09. The third-party tests they came from did not publish inspectable methodology and covered model versions OpenAI has since replaced — see the accuracy section for what replaced them.
• Session cap (4h auto-stop) confirmed via Sally.io, tldv.io, and Noota independent reviews