You recorded the client call. Now the recording sits in a folder, and the notes you promised by end of day are still a blank page. Since 6 October 2026, ChatGPT can take that audio file directly. The catch is that most people will upload it and type "summarise this", which is exactly how a wrong owner or a misheard deadline ends up in your follow-up email.
This guide shows a better way: a two-pass recording brief that separates "what was said" from "what was decided", so you can check the facts before anything goes out. You get the exact limits of the new feature, a copy-paste prompt, the mistakes to avoid with Cantonese and English mixed meetings, and a 20-minute test you can run today.
What are ChatGPT audio uploads?
ChatGPT audio uploads let paid users attach a recording to a chat and ask for a transcript, a summary, structured notes or answers about what was said. OpenAI added the feature on 6 October 2026. It works with paid subscriptions and workspaces, not the Free plan, and accepts audio files up to 512 MB.
According to OpenAI's ChatGPT release notes, you attach a supported audio file and describe what you need. OpenAI's own examples are turning a meeting, interview or lecture into structured notes or a follow-up email draft.
The practical limits, as summarised from OpenAI's help centre by DigitBin on 7 October:
--- Plans: paid ChatGPT subscriptions and workspaces. Free is excluded.
--- File size: up to 512 MB per file.
--- Formats: WAV, MP3/MPEG, OGG/OGA, FLAC, AAC, M4A, PCM, plus audio-only WebM and MP4. Files detected as video are not supported.
--- Upload cap: 80 files every 3 hours.
--- Long recordings: may be processed in sections, and very long ones may time out.
--- Accuracy: OpenAI warns transcripts may contain errors, speaker identification may be unreliable and performance varies across languages.
Availability can also vary by region, workspace settings, app version and model, so check that the paperclip accepts an audio file in your own account before you plan a workflow around it.
Why does "summarise this recording" produce unreliable notes?
A one-step summary asks the model to listen, interpret and compress at the same time. When a name is misheard or two speakers overlap, the error is silently absorbed into a confident sentence. Without a transcript to check against, you cannot see where the summary guessed, so mistakes travel straight into your follow-up.
Think about a 40-minute supplier call where someone says "let's aim for the twelfth, but I need to check with finance". A compressed summary can easily turn that into "Deadline: 12th". The hedge disappears, and now your email commits the supplier to a date they never agreed.
The fix is the same principle that makes any AI output more reliable: separate extraction from interpretation. First get a faithful record, with the uncertain parts flagged. Then build your decisions from that record, with every claim pointing back to a timestamp you can verify in seconds. If you have read our guide to output contracts for consistent AI results, this is the audio version of the same idea.
How does the two-pass recording brief work?
The two-pass recording brief splits one request into two prompts in the same chat. Pass one asks for a faithful transcript with speaker labels, timestamps and markers for anything unclear. Pass two builds a decision log only from that transcript, with each item linked to a timestamp. A follow-up email is drafted last, from the log.
Pass one: the faithful record. Upload the file and ask for a verbatim transcript, not a cleaned-up one. Tell ChatGPT to label speakers, add a timestamp at least every minute, and mark anything it is unsure of instead of guessing. Give it the attendee names in advance so it has a chance of matching voices to people.
Pass two: the decision log. In the same conversation, ask for decisions, owners, deadlines and open questions, each with the timestamp where it was said and a confidence flag. Anything that was implied but not stated goes into a separate "not stated" list. That list is where most real follow-up work hides.
Pass three, optional: the follow-up. Only after you have checked the log do you ask for an email or action list, and you tell ChatGPT to use the checked log as its only source.
Here is the prompt for passes one and two. Paste it after attaching your recording.
Try this prompt (pass one):
Attached is a recording of a [type of meeting, e.g. weekly client call] on [date].
Attendees: [Name, role], [Name, role], [Name, role].
Languages spoken: [e.g. Cantonese and English, mixed].
Create a verbatim transcript. Rules:
1. Label each speaker with the attendee name if you are confident, otherwise "Speaker A", "Speaker B".
2. Add a timestamp [mm:ss] at the start of each speaker turn.
3. Keep the original language of each sentence. Do not translate or tidy up wording.
4. If a word or phrase is unclear, write [unclear mm:ss] instead of guessing.
5. Do not summarise yet.
After the transcript, list every [unclear] marker and every place where you were not sure who was speaking.
Then paste (pass two):
Using only the transcript above, build a decision log with four sections:
DECISIONS: what was agreed, who agreed it, timestamp.
ACTIONS: task, owner, deadline exactly as said, timestamp.
OPEN QUESTIONS: anything raised but not resolved, timestamp.
NOT STATED: owners or deadlines the conversation implied but nobody said out loud.
Mark each item High or Low confidence. Use Low if it depends on an [unclear] section or an uncertain speaker. Quote the original words for every deadline and every number.
How do you use it on a real client meeting?
Run pass one, spend five minutes checking only the Low-confidence items against the audio, then run pass two and draft the follow-up from the checked log. On a typical 30 to 45 minute call, the verification step usually covers a handful of timestamps, which is far faster than replaying the whole meeting.
Say you are an account manager who just finished a 35-minute call with a retail client about their Christmas campaign. Half the call was in Cantonese, the product names were in English, and the client's marketing head and finance lead both spoke.
Step 1: Export the audio. Zoom, Teams and Google Meet can all produce an audio-only file or a recording you can convert. If you upload a video file, ChatGPT may reject it, so use the audio-only M4A or MP3.
Step 2: Run pass one with the three attendee names. Read the list of [unclear] markers at the end, not the whole transcript.
Step 3: Jump to each flagged timestamp in your player. Correct the transcript with a short message such as "At 14:20 the budget figure is HK$80,000, not HK$18,000. At 22:05 the speaker is Karen, not Speaker B."
Step 4: Run pass two. Look hardest at the NOT STATED list. In this kind of call, it often shows that nobody actually owned the final artwork sign-off.
Step 5: Ask for the follow-up: "Draft a short email to the client from the checked decision log only. Include the two open questions as questions, not as decisions." If you prefer to speak your instructions, our guide to briefing ChatGPT by voice covers how to keep spoken prompts structured.
What are the common mistakes with ChatGPT audio uploads?
The biggest mistakes are trusting speaker labels, letting ChatGPT translate mixed-language speech in the transcript, uploading confidential calls to a personal account, and feeding in very long recordings in one go. Each one is easy to avoid once you know where the feature is weakest, which OpenAI itself has documented.
--- Trusting speaker labels. OpenAI says speaker identification may be unreliable. Treat every "who said it" as unverified until you have checked the moments that matter: decisions, money and deadlines.
--- Letting it translate during transcription. Mixed Cantonese and English speech is exactly where performance can vary. If the model translates while it transcribes, you lose the evidence. Keep the original language in pass one, then ask for an English or written-Chinese version of the decision log afterwards.
--- Uploading a two-hour workshop in one file. Long recordings may be processed in sections or time out. Split long sessions at natural breaks and run pass one per part, then ask for one combined decision log.
--- Using the wrong account. Data handling differs between consumer plans and business workspaces. Client calls belong in your company's ChatGPT workspace, under its data settings, not in a personal Plus account.
--- Skipping consent. Under Hong Kong's Personal Data (Privacy) Ordinance, a recording of identifiable people is personal data. Tell participants you are recording and how the notes will be produced, and follow your company's policy.
--- Expecting it on the Free plan. Audio uploads are paid only. If your team is mixed, decide who processes recordings rather than assuming everyone can.
How can you test the two-pass method in 20 minutes?
Pick one recent recording of about ten minutes, run the old one-step summary and the two-pass brief side by side, then compare both against what you remember was agreed. The gap between the two outputs, especially in deadlines and owners, shows whether your team needs the extra pass.
--- Minutes 0 to 3: choose a short recording with at least two speakers and one decision. Export it as M4A or MP3.
--- Minutes 3 to 7: in a new chat, upload it and type "Summarise this meeting with action items." Save the result.
--- Minutes 7 to 15: in another new chat, run pass one and pass two from the prompts above. Check the Low-confidence items against the audio.
--- Minutes 15 to 20: count the differences. How many deadlines, owners or numbers did the one-step summary state more confidently than the meeting did?
If the answer is zero, a one-step summary may be enough for low-stakes internal calls. If it is one or more, use the two-pass brief for anything a client, a supplier or your boss will read. A recording only becomes useful when you can trust what comes out of it. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team. Feature details reflect OpenAI's 6 October 2026 release notes and help centre as reported on 7 October 2026; plan availability and limits may change. Sources: OpenAI release notes, OpenAI help centre, DigitBin.
Find the AI employee that fits your repetitive work
Meeting notes are one repetitive task. Answer four quick questions in UD's free AI Employee Match to see which kind of AI employee suits your workload best. When you are ready to turn it into a workflow your team can rely on, we'll walk you through every step, from tool setup to workflow design and deployment.