Back to Blog
Product Tutorial•2026-11-16•11 min read

Can ChatGPT Transcribe Audio? Plans, Limits and Fixes

Can ChatGPT Transcribe Audio? Plans, Limits and Fixes
TL
Team Laxis
Laxis Team @ Laxis

Yes, ChatGPT can transcribe audio: on a paid plan you can attach a recording to a chat and ask for a transcript, a feature OpenAI announced in its release notes on October 6, 2026. It can't on the Free plan, and in a regular chat OpenAI caps an uploaded recording at 6 minutes when you use GPT-6 set to Instant, so it suits a voice note far better than an hour-long interview. Below is what works as of November 16, 2026, checked against OpenAI's help center and developer docs, where it breaks, and what to use instead when it does.

The short answer

  • Paid plans only. OpenAI says audio uploads work on paid ChatGPT plans and workspaces, including Enterprise, and not on Free.
  • Short files work best. In a regular chat on GPT-6 set to Instant, recordings can run up to 6 minutes; anything longer is best-effort and can time out. Files can be up to 512 MB.
  • Common formats. WAV, MP3, M4A, FLAC, AAC and OGG. MP4 and WebM files that ChatGPT identifies as video are turned away.
  • Check speakers and names. OpenAI warns that transcripts may contain errors and that speaker identification may be unreliable.
  • Developers use the API instead: gpt-transcribe, 25 MB per file, $0.0045 a minute.

How to transcribe an audio file in ChatGPT

Audio uploads live in ordinary chats. OpenAI's help article adds that availability can vary by workspace settings, region, app version and the model you pick, so if the option seems to be missing, those are the four places to look.

  1. Confirm your plan. Any paid ChatGPT plan or workspace qualifies; the Free plan doesn't. On a work account, availability also depends on how your workspace is set up.
  2. Check the file. Use WAV, MP3, M4A, FLAC, AAC or OGG, under 512 MB. For a dependable result keep it to about 6 minutes, and split anything longer between sentences.
  3. Attach it to a regular chat, the same way you'd attach a document.
  4. Ask for a transcript, not a summary. Try: "Transcribe this word for word. Start a new paragraph when the speaker changes. Write [unclear] where you can't make out a word." Being explicit matters because the same upload can just as easily be summarized or discussed.
  5. Check it against the audio. Listen again to the parts that matter: names, numbers and who said what. OpenAI itself warns that transcripts may contain errors and that speaker identification may be unreliable.
  6. If it uploads but won't transcribe, convert the file to WAV and upload it again. That's the first fix OpenAI suggests.
QuestionWhat OpenAI says
Which plans?Paid ChatGPT plans and workspaces, including Enterprise. Not the Free plan.
Which formats?WAV, MP3/MPEG, M4A, FLAC, AAC and OGG
What's refused?MP4 and WebM files identified as video
How big?Audio files up to 512 MB
How long?Up to 6 minutes in a regular chat on GPT-6 set to Instant. Longer recordings may be processed in smaller sections when Data Analysis is available; best-effort, and very long recordings may time out.
How accurate?Transcripts may contain errors, speaker identification may be unreliable, and quality can differ by language.
Why is it missing?Availability can vary by workspace settings, region, app version and selected model.

Dictation and voice mode aren't transcription

Because ChatGPT has a microphone button, plenty of people assume that's how it transcribes audio. It isn't, at least not when the recording already exists.

Dictation is the microphone in the message box. You talk, OpenAI's models turn your speech into text, and you can edit it before you send it. It's built for your own prompts, spoken live.

Voice is a spoken back-and-forth with ChatGPT. The chat keeps a written record of the conversation, but OpenAI's help center describes voice transcripts as not verbatim, meaning they may not match exactly what was said. Neither feature will take a 20-minute interview off your phone and type it out; for that you need an upload, or a tool built for files.

For developers: OpenAI's transcription API

If you write code, you aren't bound by the chat limits. OpenAI's file transcription guide now points everyone to gpt-transcribe on the /v1/audio/transcriptions endpoint:

  • Files up to 25 MB in FLAC, MP3, MP4, MPEG, MPGA, M4A, OGG, WAV or WebM. For anything bigger, OpenAI says to use a compressed format or split the file into chunks of 25 MB or less, and to avoid cutting mid-sentence, which removes context and reduces accuracy.
  • $0.0045 per minute of audio, according to the model page, which is about 27 cents for an hour of recording.
  • Hints for names and jargon. A prompt describes the recording, keywords lists terms you expect to hear and languages sets the expected languages. OpenAI calls keywords hints, not required output, and tells you to check they don't make unspoken terms appear.
  • Speaker labels and timestamps run on deprecated models. Diarization comes from gpt-4o-transcribe-diarize, structured word and segment timestamps only from whisper-1. Both are on OpenAI's deprecation list, together with gpt-4o-transcribe, gpt-4o-mini-transcribe and the translations endpoint, with a shutdown date of February 26, 2027.
  • Live audio belongs in Realtime transcription, where OpenAI recommends gpt-live-transcribe.

That fourth point matters if you're building something that has to know who spoke and when: today you'd be building on models with a published end date.

Where ChatGPT transcription breaks

Anything longer than a few minutes

The 6-minute figure is the one to plan around. Past it, ChatGPT may work through the file in sections when Data Analysis is available, but OpenAI calls that best-effort and says very long recordings can time out. Splitting the file yourself works, with the caveat OpenAI gives developers: cut in the pauses between sentences, never through one, or the words at each seam suffer.

Who said what

For a voice memo, speaker labels don't matter. For an interview or a call they're the whole point, and OpenAI's help article says speaker identification may be unreliable. If a quote is headed for a story, a report or a contract, confirm the speaker against the audio first.

Timestamps

Ask for them if you need them, then check a handful against the audio before you build anything on them. A transcript with a wrong timestamp sends you to the wrong moment, which is worse than having none.

Names, numbers and jargon

Surnames, product names and acronyms are where any speech model stumbles first, because they're the words it has heard least. Listing the names you expect in your request costs nothing, but read every one back anyway. Numbers deserve the same care: on a weak recording, fifteen and fifty are easy to confuse.

Other languages

OpenAI notes that quality can differ by language. If people switch languages mid-conversation, check those stretches with particular care.

Longer than six minutes, or more than one voice?

Laxis takes a whole meeting, interview or lecture as an upload in its web app (m4a, mp3, wav or flac, up to 200 MB) and returns a transcript with speaker labels and timestamps, plus an AI summary. A free Laxis account includes 300 minutes of transcription a month.

14-day Premium trial, no credit card Try Laxis free

Fixing a file ChatGPT won't take

A failed upload is usually a file problem rather than a transcription problem. Three fixes, all possible with any audio converter or with ffmpeg:

  • A video file. Pull the soundtrack out as an MP3, then upload the MP3.
  • An M4A that uploads but won't transcribe. OpenAI's own tip is to convert to WAV and try again. Re-encoding it as an MP3 is another way to get a clean, standard file.
  • A huge WAV. Converting it to MP3 shrinks it to a fraction of the size, which matters most for the API's 25 MB ceiling.

If the recording is in the Voice Memos app, our guide to transcribing voice memos on iPhone covers the options there, and if you're starting from a video, there's our walkthrough on turning video into text.

Before you upload: privacy settings to check

What happens to a recording after you upload it depends on your plan and your settings. Here is how OpenAI describes it on its policy page about using data to improve models, last updated March 13, 2026:

  • Personal accounts. For services for individuals, such as ChatGPT, OpenAI says it may use your content to train its models. You can opt out in Data Controls or through OpenAI's privacy portal, and OpenAI says new conversations aren't used for training after that.
  • Temporary Chat. OpenAI says these chats don't appear in your history, don't use or create memories, and aren't used to train its models.
  • Business products. OpenAI says that by default it doesn't train on inputs or outputs from its business products, including ChatGPT Enterprise and the API.

Two details from OpenAI's help center are worth adding. On personal Free, Plus and Pro accounts, the Improve the model for everyone setting starts switched on. And spoken input has its own switch, Include your audio recordings, which covers dictation and voice chats. That's OpenAI's account of its own practices, not ours, so read the pages yourself before uploading anything confidential. For comparison, Laxis's commitment is written in our No-Training Pledge: "Laxis will never use your data to train AI models without your consent."

One check no setting covers: if other people are on the recording, they should have known it was being made. Some places only need one participant's consent, others need everyone's; our consent guide covers both, state by state and abroad. It's general information, not legal advice.

When ChatGPT is the wrong tool

ChatGPT makes sense for a short clip you want to transcribe and then ask questions about in the same window. For anything long, anything with several speakers, or anything you'll want to search next month, a tool built for files does the job with fewer workarounds.

  • Laxis audio to text. Upload a meeting, interview, lecture or voice memo and get the transcript back with speakers and timestamps, plus a summary. It runs on a free Laxis account with 300 minutes a month.
  • What's already on your devices. Our roundup of free ways to transcribe audio to text compares built-in options with free web transcribers and explains what really drives accuracy.
  • Meetings that haven't happened yet. Recording a meeting and having ChatGPT write it up is a different job from transcribing a file. Our guide to ChatGPT meeting notes covers OpenAI's meeting features and the prompts that hold up.

Where Laxis fits

Laxis is an AI meeting assistant that also takes files: upload a recording in the web app and the transcript comes back with speakers and timestamps, alongside a summary and action items. The same account covers live conversations. The desktop app records Zoom, Google Meet, Teams and Webex calls through your computer's own audio, with no bot joining by default, and the phone app handles in-person meetings and phone calls. The Laxis meeting assistant page has the full picture.

The limits, plainly. In the standard apps, your audio is processed on Laxis's cloud servers. Basic, the free plan, allows 300 minutes a month, and meetings and dictation both count against them; a single recording can run 45 minutes, and anything more than 30 days old moves to an archive (kept, not erased) until you upgrade. Transcription covers 60+ languages. Laxis is not HIPAA compliant, and it signs no BAAs. And when you capture a call with no bot in it, let the other people know: the app prompts you before you record, and it takes one sentence.

And if all you have is a two-minute voice note and a ChatGPT subscription, use ChatGPT. Pick the tool that fits the file.

The bottom line

ChatGPT finally accepts audio files, and for short clips on a paid plan it's genuinely convenient: transcribe, then ask about what was said, without leaving the chat. The catches are the 6-minute ceiling on the Instant setting, speaker labels you can't rely on, and a Free plan that can't do it at all. For long recordings with several voices, start with a tool built for transcription, then bring the text to ChatGPT if you still want to talk it through.

Transcribe the whole recording, not the first six minutes. Laxis turns meetings, interviews and voice memos into summaries, action items and transcripts with speakers. 14-day Premium trial, no credit card. Download Laxis

Frequently asked questions

Can ChatGPT transcribe audio for free?

No, ChatGPT cannot transcribe an uploaded audio file on the Free plan. OpenAI says audio uploads are available with paid ChatGPT plans and workspaces, including Enterprise, but not on Free. Dictation and voice mode transcribe speech as you talk, not a recording you already have, so they are not a workaround. For free file transcription, use a dedicated transcription tool.

What audio formats can ChatGPT transcribe?

ChatGPT can transcribe common audio formats, including WAV, MP3, M4A, FLAC, AAC and OGG, according to OpenAI's help center. MP4 and WebM files that are identified as video are rejected by audio uploads, so extract the audio track first. If a file uploads but will not transcribe, OpenAI suggests converting it to WAV and uploading it again.

How long can an audio file be for ChatGPT to transcribe it?

OpenAI says a recording uploaded to a regular chat can be up to 6 minutes long when you use GPT-6 set to Instant, and audio files can be up to 512 MB. Longer recordings may be processed in sections when Data Analysis is available, but OpenAI calls this best-effort and warns that very long recordings may time out.

Can ChatGPT identify different speakers in a recording?

ChatGPT cannot reliably identify different speakers in a recording. OpenAI's help center warns that transcripts from audio uploads may contain errors and that speaker identification may be unreliable, so check who said what against the audio. Developers can get speaker labels from the API model gpt-4o-transcribe-diarize, but OpenAI has deprecated it and plans to shut it down on February 26, 2027.

Does OpenAI use uploaded audio to train its models?

It depends on your plan and settings, according to OpenAI. For individual services such as ChatGPT, OpenAI says it may use your content to train its models unless you opt out in Data Controls or its privacy portal, and Temporary Chats are not used for training. OpenAI also says that by default it does not train on inputs or outputs from business products such as ChatGPT Enterprise and the API.

How much does the OpenAI transcription API cost?

OpenAI's recommended transcription model, gpt-transcribe, costs $0.0045 per minute of audio, or about 27 cents an hour, according to its model page. Files sent to the transcriptions endpoint can be up to 25 MB, in formats including MP3, MP4, M4A, WAV, WebM, FLAC and OGG. Larger recordings need compressing or splitting, ideally between sentences rather than in the middle of one.

Related Product

Laxis Meeting Assistant

Automate your notes today.

  • Bot-free notes for Zoom, Google Meet, Teams and in-person meetings
  • AI summaries, action items and follow-ups in 60+ languages
  • Mac, Windows, iPhone and Android, with a free plan

Start for free, no credit card required.

Learn More ▷