You can attach a recording file to ChatGPT and ask for a transcript, a summary, or a draft of meeting minutes. This feature is available on supported paid plans. It works differently from Voice, which lets you talk with ChatGPT, and from Record, which captures audio as it happens. If you want to use an existing MP3, M4A, or similar file, start by attaching the file.
3 steps to turn a recording into minutes
Give ChatGPT the audio file
Check the supported formats and size
Start with a transcript in the original language
Leave unclear parts marked
Decisions, owners, and deadlines
Check important parts against the recording
The limit is 512MB. Long recordings can stop partway through, even when they are within the size limit.
We checked these specifications against OpenAI's official English help article on October 7, 2026. Source: Uploading audio files, supported formats, and limits.
Contents
- 1. Available plans and the differences between four audio features
- 2. Supported formats, size, and long-recording limits
- 3. Steps to attach audio and transcribe it
- 4. Prompts for meeting minutes, interviews, and lectures
- 5. How to check speakers, numbers, and missing parts in a transcript
- 6. Pricing, usage limits, and differences from the API
- 7. When uploads fail or the file can't be read
- 8. How to think about training use, storage, and deletion of recordings
- Summary: Attach the recording, check the transcript, then use it
- FAQ
1. Available plans and the differences between four audio features
OpenAI's official help says that audio upload is available on paid ChatGPT plans and in workspaces, and includes Enterprise. Free is not included. Being able to attach documents or images for free does not mean that audio attachments follow the same rules. Depending on your region, workspace settings, app or browser version, and the model you select, what you see may vary.
Saved recording → Audio upload
Attach an MP3, M4A, or similar file to a conversation. Use it for transcripts and questions about what was recorded. This article covers this feature.
Want to talk → Voice
A feature for talking with ChatGPT by voice. A voice conversation record is different from a full transcript of a recording file.
Speak a request → Voice dictation
Turns a request you speak into the microphone into text, which you can edit and send. This is separate from choosing a recording file.
Record on the spot → Record
Records and organizes meetings or voice memos. The official Record guide is for the macOS app. Check recording length limits and how recordings are saved separately from file attachments.
Sources: Voice, Voice dictation, Record. To learn about the models and mechanics behind voice conversation, see our GPT-Live explainer.
Being able to attach a meeting recording and having a tool join a meeting automatically are two different things. To compare recording methods and dedicated services, see our guide to AI meeting minutes tools and how to automate them.
2. Supported formats, size, and long-recording limits
| Item to check | What the official documentation says |
|---|---|
| Supported audio | WAV, MP3/MPEG, OGG/OGA, PCM, FLAC, AAC, M4A, and audio-only WebM or MP4. |
| File size | Audio files up to 512MB. |
| Files that include video | WebM or MP4 files recognized as video are not supported for this audio upload. |
| Recording length | Long recordings may be split for processing when Data Analysis is available. Very long recordings may stop processing because of a timeout. |
| Accuracy | Accuracy varies by language. The transcript may contain errors, and speakers may not always be told apart correctly. |
Source: OpenAI's audio upload specifications. The file itself must contain readable audio data. Changing only the file extension does not change the actual format or whether the file contains video.
Before you attach a file, check its size in the file properties or information screen.
Even within the size limit, processing of long recordings may time out. The official help does not state a recording length that is guaranteed to process.
If a long file fails, try splitting it into shorter parts on your own computer and attaching them in order. Record the order of the parts ("first half," "second half"), where each part starts in the original recording, and any short overlap between parts. This makes it easier to check later for gaps and duplicates. If you use an outside service to split the file, confirm that you are allowed to send a recording that contains confidential information.
3. Steps to attach audio and transcribe it
- Check the file. Play the beginning, middle, and end in a local player to confirm that there is sound throughout. Note the format, size, and recording length.
- Open a conversation on an account with a supported paid plan. Choose the supported audio file from the attach option near the message box. Button names and placement depend on the app version, so don't confuse the attach button with the microphone for voice dictation.
- Send your request together with the attachment. Start by asking for a "transcript." If you ask only for a short summary at first, you'll have less information to compare against the recording.
- Check the result, then make follow-up requests in the same conversation. Correct any errors first, then turn the transcript into minutes or a draft email. Also check whether the response covers the entire recording through to the end.
The official basic workflow is: attach the file, state the processing you want, then check the response and make follow-up requests. Here is an example prompt that follows those steps.
Transcribe the attached audio in its original language.
In the first output, do not summarize or translate. Keep the content and order of what was said.
Mark unclear parts as [unclear] and do not fill them in with guesses.
Add speaker labels only when you can reliably tell who is speaking.
If you cannot identify the speaker, write [speaker unknown].
Also indicate any sections you could not process, and any points where you cannot confirm that the whole recording was processed.
Recording language: English
Recording length: (fill in if known)
Giving the recording length helps you check which part of the recording the output covers. If you add specialized terms for spelling, also say "use them only when they are actually spoken," so the model doesn't use them to fill in statements that were never made.
4. Prompts for meeting minutes, interviews, and lectures
After you have checked the transcript, change the output to suit your purpose. The key is to separate what can be established from the recording from what the AI suggests. The examples below are prompts for each use case.
Meetings: separate decisions, proposals, and open items
Based on the transcript we just checked, please create meeting minutes.
Output: meeting purpose / confirmed decisions / open issues / next actions.
Present the next actions in a table with "content, owner, deadline, and the statement it is based on."
Do not turn proposals into decisions.
If the owner or deadline is not mentioned, write "TBD."
If you add AI suggestions, put them under a heading separate from the facts in the recording.
Example: "It would be good to be able to check this by Friday."
Don't write "The deadline is Friday and Alex is responsible." Don't add an owner who was not mentioned.
Record it as: "A check by Friday was proposed. Owner and deadline to be confirmed."
Interviews: separate quotes from readable prose
Organize the confirmed transcript by topic.
Clearly separate "quoted statements" from "summaries of content."
Do not rewrite the wording of quotes to make them read more smoothly.
Use a speaker's name only when the recording or information I have given you identifies them.
Also list the proper nouns, numbers, and unclear parts that should be checked again before the article is used.
Check the recording again for any passage you plan to publish as a quote. Light editing can make a quote more readable, but changing an ending, a hedge, or a negation can change its meaning. A label such as "Speaker A" does not by itself tell you who said something.
Lectures: separate review notes from added explanations
Create review notes from the confirmed transcript.
Divide them into "content explained in the lecture," "terms," and "questions to check understanding."
If you add explanations that are not in the recording, label them "Added explanation."
Mark formulas, names, and technical terms that are unclear in the audio as items to verify, rather than stating them as facts.
First create the notes in the original language, and do not translate yet.
If you want to translate into another language, correct the original transcript first and then request the translation. Translating a transcript that still contains recognition errors makes the places that need fixing harder to spot. If your goal is subtitles or video editing, see our guide to choosing between subtitle and transcription tools. If you ask ChatGPT for SRT format, check the timecodes against the video.
5. How to check speakers, numbers, and missing parts in a transcript
OpenAI also lists transcription errors and uncertain speaker identification as things to watch for. Don't finish your verification by simply asking the AI "Is there anything wrong?" Compare the output with the original audio. Even when a summary reads well, check the following against the recording.
Dates, amounts, headcounts, and percentages. Check numbers and units that sound similar against the recording.
"Won't," "hypothetically," "still undecided." If words like these go missing, the meaning of the statement changes.
Check that speakers have not been mixed up. Distinguish proposals from confirmed commitments.
Check that topics from the beginning, middle, and end are all covered. Confirm that processing did not stop partway and that no sections are missing.
Checking the beginning, middle, and end is a quick way to catch large omissions. For important uses such as contracts or delivery dates, listen again to each relevant statement. Before using the transcript as an official record, also have the people involved review the content.
If you find more unclear parts in sections with background noise or overlapping speech, first listen to those sections on your own. Repeatedly asking the AI to rewrite its answer can leave errors in place while making the text read more smoothly.
6. Pricing, usage limits, and differences from the API
Attaching audio to a ChatGPT conversation and calling the developer API are billed and limited separately. The attachment feature does not use an API key. If you build transcription into your own app through the API, that is API usage, which is separate from a ChatGPT subscription.
| Comparison item | Attaching audio in ChatGPT | Transcription API |
|---|---|---|
| How it's used | Attach a file to a conversation and make a request. | Send audio from a program. |
| Plan and billing | Requires a supported paid ChatGPT plan or workspace. | The API is billed separately from the ChatGPT monthly subscription. |
| Size limit | Audio up to 512MB. | The File transcription guidance lists 25MB. |
| Where to check specifications | ChatGPT audio upload help. | API File transcription guide and pricing page. |
Sources: Separate billing for ChatGPT and the API, File size limit for the transcription API. Don't present the API's 25MB limit as the attachment limit for ChatGPT, and don't assume that a Pro plan includes API charges.
Pricing and usage limits for audio attachments
The official audio upload help does not list extra per-minute charges for individual users, or processing times and counts specific to audio for each plan. If you reach a usage limit or see an error, check the guidance for your own plan and workspace.
7. When uploads fail or the file can't be read
"I can't select the file," "the file was attached but can't be processed," and "the result stops partway" each have different things to check. Narrow down the cause by symptom before changing your plan. Here is the order of checks, based on the official requirements.
| Symptom | Check first | Next step |
|---|---|---|
| Can't attach audio | Whether you are on a paid plan that supports audio upload, and whether the format and size are supported. | Also check the region, model, app version, and organization settings. Being able to attach a document does not mean you can attach audio. |
| Told the file can't be read after attaching | Whether you can play the audio locally. Whether the file is an MP4 or WebM that is recognized as video. | Narrow it down with a short, correctly exported audio file. Don't try to change the audio format just by renaming the extension. |
| Long recording stops or content is missing | How far the response covered, and whether the end of the recording is included. | Consider splitting the original recording into shorter parts, noting their order and start positions, and processing them again. |
| Upload limit error | Your account, plan, and storage capacity, and any OpenAI outage notices. | Don't retry many times in a short period. OpenAI notes that failed uploads may still count toward the limit. |
Source: OpenAI's official upload troubleshooting. If the problem persists, note the date and time with the time zone, the device and browser, the selected model, the format, size, and recording length, the error screen, and the request ID if you can get one, so you can include them in a support request. In screenshots, avoid showing more of the recording's content or personal information than necessary.
For a short test, use an audio file you made yourself that contains no confidential information. Work out whether every file fails or only certain formats or lengths do. Before upgrading your plan, suspect problems with the environment or the file.
8. How to think about training use, storage, and deletion of recordings
Before you upload a recording, check the explanation and consent required from participants, and the information-handling rules at your workplace. Audio can include names, addresses, customer information, and background conversation that you did not include in your request. Decide first whether you can use material with unnecessary sections removed, and whether it is acceptable to send it to the cloud.
For consumer services, content may be used for training depending on your settings. Even on Pro, the terms for training use differ from those for business contracts.
Deleting a conversation may not delete the recording file saved in Library. Check each storage location.
According to OpenAI's general policy, consumer ChatGPT services may use content for training. Turning off "Improve the model for everyone" excludes new conversations, but there are exceptions for feedback you choose to send. Business, Enterprise, and Edu, as well as the API, do not use inputs and outputs for training by default. Source: how data is used for training, by service.
However, the audio upload help does not explain how training use of an attached audio file itself relates to the "Include audio recordings" setting in Voice. Do not apply the explanations for Voice clips or Record recordings directly to the MP3 or M4A files you attach. For guidance on reading the settings and organizing the information you have entered, see ChatGPT and Codex training use and confidential information.
If a file is saved in Library, deleting the conversation alone will not remove the copy in Library. Transcripts and summaries may remain in the conversation, so check each storage location. Archiving is not the same as deleting. Source: saving and deleting chats and files in ChatGPT. Also keep in mind that turning off training does not stop a recording from being sent off your device.
Summary: Attach the recording, check the transcript, then use it
To use a saved recording in ChatGPT, attach the audio file to a conversation and first ask for a transcript. A practical sequence is to check decisions and quotes against the recording, then organize the material into minutes, an article, or review notes.
- Requirements: Audio upload is for paid plans. It is not available on Free.
- Material: Check the supported audio formats and the 512MB limit. Distinguish MP4 and WebM files that include video.
- Quality: Check speakers, numbers, and negations against the recording. Also confirm that the transcript covers the entire recording.
- Cost and data: The API is billed separately from ChatGPT. Pricing and usage limits specific to audio attachments are not published in full, and conversations and files may be stored separately.
This article explains procedures based on the specifications in OpenAI's official help. We did not measure transcription accuracy or processing time with actual recordings.
FAQ
Can I transcribe an MP3 with the free ChatGPT plan?
According to the official help, audio file upload is for paid plans and workspaces, and Free is not included. Being able to use free voice conversation is a separate matter.
Can I also attach M4A and MP4 files?
M4A is one of the supported formats. MP4 and WebM are supported only when they contain audio alone. Files recognized as video are not supported for this audio upload. The file must contain readable audio data.
If a multi-hour recording is under 512MB, will it work?
You can't tell from the file size alone. The official documentation says that long recordings may be split for processing, and that very long recordings may time out. The size limit and the processable length, as well as the accuracy of the full transcript, are separate questions.
If I have ChatGPT Pro, is the transcription API included in the price?
No. API usage is billed separately from a ChatGPT subscription. Keep attaching audio to a conversation separate from calling the API from your own program.