Automatic meeting minutes from a recording
Automatic meeting minutes are only useful if the person who missed the meeting can read them in two minutes and know what was decided and what they owe. This is how to get from a recording to that page: what to capture in the room, which settings to choose, and what to check by hand before you press send.
Transcribe a meeting or write the minutes: two different jobs
When people say they want to transcribe a meeting, most of the time they want something else. A transcript is everything that was said, in order: the false starts, the five minutes about the broken projector, the same point made three times by three people. An hour of talk comes out at roughly forty pages. It is a record, and a good one to keep, but nobody who skipped the meeting is going to read it.
Minutes are what has to be remembered. They answer four questions: what was this about, what did we decide, who does what by when, and what is still open. If the page answers those, it has done its job, and it can do it in a single side of A4.
The two documents are worth having together, because they cover for each other:
- The minutes are what you send. They are short enough to be read on a phone between two other things.
- The transcript is what you keep. When someone says "that is not what we agreed", you go back to the exact words.
If the transcript is all you need (for subtitles, for quoting someone exactly, for a searchable archive), use the tool built for that: you can transcribe a video to text and get TXT, SRT, VTT and JSON files in one ZIP, with an option to tell speakers apart. For minutes, you need the step after transcription: the summary.
What useful minutes contain, and why timestamps matter
Good minutes have the same five parts whether a person writes them or a model does. Use this as the checklist for whatever comes out of the machine:
- Subject. One line: which meeting, which date, what it was for. "Weekly sync" tells nobody anything in three months. "Q1 budget review, 12 March" does.
- Decisions. Written as facts, not as discussion. "The launch moves to 4 May" rather than "we talked about the launch date".
- Tasks with an owner. A task without a name on it is a wish. Each one needs a person and, when it was said, a date.
- Open points. What was raised and not closed. This is the part people leave out and the part that saves the next meeting from starting at zero.
- Timestamps. The minute of the recording where each topic starts.
The last item is the one that turns a summary into something you can trust. A summary compresses, and compression loses nuance. If the line reads "14:20 Pricing: agreed to keep the current tiers", anyone who doubts it can open the recording at 14:20 and hear the two minutes that matter, instead of scrubbing through an hour. Minutes without timestamps ask the reader to believe them. Minutes with timestamps let the reader check.
Notice what is not on the list: who said what at every turn, the order of the conversation, and the reasoning behind each position. That lives in the transcript. Minutes that try to carry it stop being read.
Record the meeting so the summary comes out right
The quality of the audio decides nearly all of the result. A model cannot summarise a sentence it did not hear properly, and it will not tell you it guessed. Three habits make more difference than any setting you choose afterwards:
- Put a microphone near whoever is speaking. A laptop at the far end of a six-metre table records the air conditioning and the person sitting next to it. In a room, put a phone or a conference microphone in the middle of the table, no more than about a metre from each voice. On a video call, every participant on their own headset beats one shared room microphone.
- Ask people not to talk over each other. Two voices at once become one garbled line in the transcript, and the decision that got made in that overlap is the one that goes missing. It feels stiff to say at the start of a meeting. It takes ten seconds and you only have to say it once.
- Say your name before you speak. "Priya here: I can take the supplier call." This sounds formal, but it puts the owner of the task into the audio itself, so the text carries it without depending on voice recognition. It matters most for exactly the lines you care about: commitments.
Two smaller things help. Say dates and figures in full ("the fourteenth of March", "twelve thousand five hundred") rather than "next Thursday" or "twelve five", because the written version will be read weeks later by someone without the context. And close the meeting by reading the decisions and tasks aloud. That thirty-second recap is the cleanest source material the summary will get: one voice, no interruptions, the whole outcome in order.
If your call app only exports video, that is fine: you do not need to extract the audio first. If it exports audio only, that works too.
Summarise a recorded meeting, then check these four things
When you summarise a recorded meeting automatically, the errors are not spread evenly. They cluster in the places where a mistake costs the most. Read the draft against the recording for these four, every time, before it leaves your outbox:
- Proper names. People, clients, products, places. Speech recognition writes what it hears, and an unusual surname or a brand name is often heard as a common word. Check every name against the invite list.
- Figures. Budgets, quantities, percentages. "Fifteen" and "fifty" are close enough in sound to swap, and a wrong figure in minutes becomes a wrong figure in somebody's spreadsheet.
- Dates and deadlines. Especially relative ones. "By the end of next week" needs turning into a real date, and you are the one who knows which week that was.
- Assigned tasks. This is the honest limit of the method. A summary works from text, and when several people discuss who should take a job, the model can attach the task to the person who talked about it rather than the person who accepted it. It can also record a suggestion ("maybe we should ask legal") as a decision. Use the timestamp, listen to that minute, and correct the owner.
That is also where the usual advice goes wrong. "Record it and let the AI send the minutes" skips the one step that makes minutes safe to act on. Treat the automatic document as a first draft written by a careful intern who was not in the room: complete, well ordered, and in need of ten minutes from someone who was.
There are also meetings where this is the wrong tool altogether. If the audio is a phone on a café table with four people talking at once, no summary will rescue it: write the minutes from memory while it is fresh. And for anything with legal weight, such as a board resolution or a disciplinary hearing, the automatic draft can save typing, but the wording has to be approved by the people in the room. Finally, ask before you record. Everyone present should know the meeting is being recorded and why.
Turn a meeting recording into a document in Airpic, step by step
The tool for this is Airpic's video to document converter. It takes video or audio (MP4, MOV, WebM, MKV, MP3, WAV, M4A) up to 3 hours and 1 GB, with no account. These are the settings for minutes, and the reason for each:
- Upload the recording. Drop the file exported from your call app or your phone. A card appears with the file name and a frame from the video, so you can confirm it is the right meeting.
- "I want": choose "Summary + full transcript". You get the page to send and the record to keep in one file. Pick "Summary only" if the document goes to people outside the team and the raw conversation should not travel with it.
- "Template": choose "Meeting (decisions and tasks)". The general template writes a summary and key points. The meeting template adds a section for what was decided and who does what, which is the part your readers look for first.
- "Summary length": choose "normal (2-4 paragraphs)". For a one-hour meeting, this is the version that gets read in two minutes. "Detailed (in sections)" suits a long workshop with several separate topics. "Short (one paragraph)" is for a stand-up.
- "Document language". Leave it as "same as the audio", or pick another language if the minutes go to a team that works in one. The summary, key points and chapters are written in the language you choose. The transcript stays in the language that was spoken.
- "Names and vocabulary". Type the names of the attendees, the client and any product or project names, separated by commas. They are used as a hint during transcription and near-miss spellings are corrected afterwards. This one field removes most of the name errors from the checklist above.
- "Format". DOCX if you are going to correct it, which for minutes you are: it opens in Word or Google Docs. PDF once it is final and ready to circulate. Markdown if your team keeps notes in a wiki or a notes app.
- Process and read the preview. A Markdown preview of the document appears on the page: title, summary, key points with their minute, chapters with timestamps and, if you asked for it, the transcript in timestamped paragraphs. Read the tasks first.
On timing, a measurement of our own: in our tests on 4 September 2026, a 3-minute interview became a PDF with a detailed summary in 63 seconds, and the transcript alone, as Markdown, took 17 seconds. A full meeting takes longer in proportion, and the first job of the day can add a minute or two while the engine starts. The progress bar gives an estimate.
Processing and seeing the result is free. You get one free full download a month, and after that downloads use credits that do not expire, with no subscription. The recording is deleted 1 hour after you download, 1 day at most, it is never used for training, and the GPUs that process it are in the European Union, which is usually the first question when the meeting was an internal one.
If the same meeting had a moment worth sharing more widely, such as a product demo or a short announcement, the same recording can go through the tool to make shorts from a long video, with word-by-word captions. And if the final PDF needs sign-off from the attendees, you can sign a PDF on your own computer with Airpic PDF. Start with the minutes: turn your meeting recording into a document and have the draft in front of you before the next meeting begins.
FAQ
Can AI write meeting minutes from a recording?
Yes. The recording is transcribed with timestamps and a summary is written from that text: subject, key points with their minute, decisions and tasks. Treat it as a first draft. Check names, figures, dates and who owns each task against the recording before you send it.
What is the difference between a meeting transcript and meeting minutes?
A transcript is everything that was said, word for word, and an hour of talk runs to dozens of pages. Minutes are the short version that matters afterwards: what was decided, who does what, and what is still open. Send the minutes and keep the transcript as the record.
How do I get a summary of a recorded video call for free?
Export the recording from your call app as video or audio and upload it to Airpic's video to document tool. Processing and viewing the result is free with no account, and you get one free full download a month as DOCX, PDF or Markdown.
Is it safe to upload a confidential meeting recording?
On Airpic the file is deleted 1 hour after you download the result, 1 day at most, it is never used for training and it is processed on GPUs in the European Union. Whatever tool you use, tell the participants the meeting is being recorded and follow your organisation's policy.
We measured it
- A 3-minute interview turned into a PDF with a detailed summary in 63 seconds; transcript only in Markdown, in 17 seconds. (04-09-2026)
Measured on airpic.app with the GPU already running. If nobody has used it for a while, the first job takes 40 to 90 seconds longer to start.
Turn your meeting recording into minutes
Keep reading
AI video editor: from ten phone clips to a reel with captions
How an AI video editor turns ten loose phone clips into a 30-60 second reel with automatic captions: what to keep, pace, colour, 9:16 and a 1080p export.
Text to video AI: write a prompt the generator understands
Text to video AI guide: how to write a prompt an AI video generator understands (shot, subject, action, light, camera), pick the format and test with drafts.