Skip to content

AI meeting minutes

AI meeting minutes: how a meeting becomes minutes

AI meeting minutes are a summary of a meeting that is written from the recording or from the transcript, without anyone taking notes. Three steps are behind it: what was said is turned into text, the speakers are told apart, and the text becomes a draft with topics, decisions and action items.

The second step decides whether the result is usable. Minutes in which “Speaker 2” promises something are not a list of action items. Tools that receive a finished mixed recording have to guess who is speaking. Raven does not receive a mixed recording: it records one audio track per participant, and the name is attached to the track.

The third point is the language. In German, the quality of the speech recognition decides everything that follows. So we claim nothing there and publish the measured error rates per dataset.

Start free

The app is in German for now.

See the measured error rates
01How it works

Three steps turn the conversation into a draft

  1. First: speech becomes text.

    The recording goes through speech recognition, which turns what was said into running text with timestamps. This is where most errors arise, and where it is decided whether a technical term, a surname or a number arrives correctly.

  2. Second: the text gets its speakers.

    Every line is assigned to the person who said it. Without this step no model can say later who took on a task.

  3. Third: the text becomes a draft.

    A language model reads the transcript and structures it: topics discussed, decisions made, action items with the person responsible and the deadline, open points. What comes out is a draft, not a finished document. Reading it through remains your job, and you do it before you send it.

02Speakers

Separate audio tracks beat any guess

The same moment in the conversation: on the left a cut that someone has to decide on. On the right two lanes that never touch.

There are two ways to get names into the minutes.

The usual way: a tool receives a finished recording in which all voices lie on top of each other, and divides it into segments by the characteristics of each voice. The result is called “Speaker 1” and “Speaker 2” at first, and someone has to add the names. With two calm voices that works well. With five people talking over each other, with poor microphones and with two similar voices, it is a guess.

The way in Raven: the meeting takes place in Raven, so every person is recorded separately, one audio track per participant. The name is attached to the track, not to an assumption about the voice. During transcription every track is read on its own and then put together in order. People who talk at the same time still get two clean lines.

You correct misspelled names in the transcript. The change then applies to every place in the meeting, not just the one where you made it.

The transcript with namesA transcript in Raven: every line with a timestamp and the name of the person speaking, two overlapping contributions shown as separate lines, and one speaker name being corrected.00:12Anke00:31JanosJonas00:31Miriam01:04Tarek01:20JanosJonas
Two contributions at the same time stay two lines. And a name you correct once is correct everywhere.
03German

German is not a setting, it is something you measure

“Supports German” is on every product page in the field. It only says that a box can be ticked, and nothing about how many words are wrong afterwards.

We do not claim to use the best model. We measure the word error rate for German audio, publish it per dataset and release the tool the measurement is made with. So you can check the numbers instead of taking our word for it.

Speech recognition runs mostly on our own machines in Germany, with a model specialized in German. If that hardware fails, a rented machine in Frankfurt takes over; on the Souverän plan, a third-party attested enclave in France. All of it within the EU. No US provider is in the data path of meeting content.

04Comparison

Which AI writes meeting minutes?

Five tools, four questions. Every statement about another product comes from that product's own page, retrieved on September 3, 2026, with the source under the table. Where the page checked does not state something, the cell says so.

Five tools, four questions
CriterionWhere speaker separation comes fromGermanWhere meeting content is processedWhat it costs
RavenSeparate audio tracks, one per participant. The name is attached to the track.A model specialized in German, error rates published per datasetGermany and the EU. No US provider is in the data path of meeting content.Free plan: 5 meetings a month, up to 4 people, up to 45 minutes each. Paid from the Solo plan, amounts on the pricing page.
KlangSource [1]Recognizes the speakers automatically from the recordingTranscription in German, by its own accountAll data in Europe, by its own accountFree plan without a credit card, by its own account. The page checked names no amounts.
FirefliesSource [2]A notetaker joins the meeting. The page checked does not describe how speakers are separated afterwards.More than 100 languages, by its own account. The page checked names no measured quality for German.SOC 2 Type II, GDPR and HIPAA are named. The page checked names no processing location.Free plan. Pro 10, Business 19, Enterprise 39 US dollars per seat per month, billed annually.
OtterSource [3]Speakers are named. The page checked does not describe how they are separated.German is one of six transcription languages, by its own accountTLS and AES-256 are named. The page checked names no processing location.Free plan with 300 minutes a month and 30 minutes per conversation. Pro 8.33, Business 19.99 US dollars per user per month on an annual subscription.
ChatGPT, record modeSource [4]Separates several speakers and assigns labels such as “Speaker 1”; you add the names afterwards.Works best in English by its own account, other languages vary.On Plus and Pro, transcripts may be used for model training as long as “Improve the model for everyone” is switched on. Business, Enterprise and Edu are exempt.Only in the macOS app and only on the Plus, Pro, Business, Enterprise and Edu plans.
  • [1]klang.ai/de/ki-meeting-protokoll/, retrieved September 3, 2026
  • [2]fireflies.ai/pricing and fireflies.ai/security, retrieved September 3, 2026
  • [3]otter.ai/pricing, retrieved September 3, 2026
  • [4]OpenAI Help Center, “ChatGPT Record”, retrieved September 3, 2026
  • Statements about Raven from its own source code, as of September 3, 2026. Other providers' prices in US dollars, because that is how they state them.

Can ChatGPT write meeting minutes?

Yes. In the record mode of its macOS app, ChatGPT records a meeting, takes notes and summarizes it. For minutes you use at work, though, three other points decide.

  1. With an ordinary account, the data goes to the United States.

    For users in Europe, OpenAI states in its privacy policy that it also processes and stores personal data on servers in the United States. On Free, Plus and Pro, transcripts from record mode may also be used to train its models as long as “Improve the model for everyone” is switched on.

  2. A European storage location does not change where the company sits.

    OpenAI offers storage in Europe for the API and for the business editions ChatGPT Enterprise and Edu. It remains a US company based in San Francisco and is subject to the US CLOUD Act (18 U.S.C. § 2713). The statute obliges US providers to hand over, when ordered, data in their possession or control, even if that data is located outside the United States.

  3. ChatGPT is not specialized in German.

    According to OpenAI's own help article, record mode works best in English today. Accuracy for other languages is improving but can vary. Raven is measured on German: on German read-speech datasets its word error rate is 5.8%, as of November 9, 2026. All error rates per dataset.

Where the recording ends up and how well the tool understands German are therefore the two questions to settle before the first set of minutes. Raven processes meetings with no US provider in the data path of meeting content and measures its speech recognition on German.

Sources

  1. OpenAI: Europe privacy policy, section 10 “Data transfers”Version of August 24, 2026, read through the Internet Archive on September 26, 2026 · Archived copy
  2. OpenAI: Introducing data residency in EuropeVersion of February 5, 2025, read through the Internet Archive on September 26, 2026 · Archived copy
  3. OpenAI Help Center: ChatGPT RecordRead on September 26, 2026
  4. Legal Information Institute: 18 U.S. Code § 2713 (CLOUD Act)Read on September 26, 2026
05Consent

No recording without consent

AI meeting minutes require a recording, and a recording as a rule requires the consent of everyone taking part. That is not a formality to be supplied later.

In practice: say at the start that the meeting is being recorded and what for. Ask whether anyone objects. Note in the minutes that you asked. And do not let anyone be recorded who was not in the room when you asked.

That is why no recording starts in Raven without someone having decided it. What Raven does by itself afterwards, and what only after your click, is stated at each feature.

06Raven

What Raven does, and what it does only after your decision

Raven is a meeting platform from Germany. The meeting takes place in Raven, is recorded and transcribed. No additional participant sits in the call to do that, because the call already runs here.

The summary follows every recording by itself. As soon as the transcript is ready, Raven writes it without you having to request it. If you do not want that, you switch it off once in the settings. Then it is written only when you request it. The decision that counts is made beforehand: whether to record.

The draft has four blocks: topics discussed, decisions made, action items with the person responsible and the deadline, open points. Every decision is assigned to the person who, according to the transcript, said it. The model may not add numbers, deadlines or names that were not mentioned, and an open point stays open instead of becoming a result. You can change the instruction the draft is written by, for yourself alone or for your team. How minutes are structured and which wording belongs in them is described in writing meeting minutes.

The minutes leave Raven as a PDF or as Markdown, with only the summary, only the transcript or both, and by email straight from the meeting. If your committee requires a fixed form, there is the template for meeting minutes to download.

Recording, speaker attribution and summary are already included in the free plan. It covers five meetings a month with up to four people and up to 45 minutes each, without video recording. The full ladder is on the pricing page.

07Instruction

What your draft looks like is yours to prescribe

In Raven you decide yourself what your draft should look like. The field is called “Eigene Anweisung” (your own instruction) and sits in the settings under “Meetings”: an empty text field, a few sentences in your own words, no checklist and no construction kit. The next draft follows them. Why a law firm writes something different there than a sales team is explained under writing meeting minutes.

The instruction belongs to you, not to your company's account. Your team can store a shared one so that minutes look the same everywhere; your own takes precedence. If you have not written one, the team's applies, and without either the standard structure of topics discussed, decisions made, action items with the person responsible and the deadline, and open points.

Two rules stand above every instruction, yours included. Whoever said something remains the one who said it: no statement moves to another person. And what did not come up in the conversation is not in the draft: no number, no amount, no date that nobody mentioned, and what was open stays open. Both are placed before your instruction before the draft is written. They cannot be written away, and that is exactly why you can write whatever you want for the rest.

Your own instruction and the two fixed linesA slip of paper in two parts: at the top two pre-printed lines that apply to every draft and cannot be changed, namely that every statement stays with the person who made it and that nothing is invented. Below it the free field with a self-written instruction that is being edited.applies to every draftNames stay with their sentencesNothing that was not saidKeep it short, one paragraph per topic.Deadlines always with a date.Customer objections separately
The lower part is yours and is an empty field. The upper part is on every slip and cannot be overwritten.
08Connections
  • Open standardCalDAV and CardDAV

    The calendar and the address book on your own server or with your own provider, such as Nextcloud, mailbox.org or iCloud.

  • GoogleCalendar and contacts

    Two separate connections, one for the calendar and one for the contacts.

  • MicrosoftOutlook

    Calendar and contacts run over the same connection, because at the provider they belong to one product.

Your calendar, even if it is not with Google

The assistant in Raven does not work on a second copy of your appointments but on your calendar. It searches it, answers when you are free, creates appointments, moves them and cancels them. That is the same for all three connections.

None of that happens unasked. Before an appointment is created, moved or canceled, the assistant puts the proposal in front of you with date, time and participants and waits for your confirmation.

It reads your address book along the way: it finds the person for the name and the address for the person. Nothing is written there. If you maintain your contacts, you keep doing that where you always have.

The difference is in the first line of the list. If you keep your calendar with Microsoft or Google, you find a connection everywhere. If you keep it on your own server or with another provider, mostly nowhere: Zoom's own documentation names three providers and no open standard (retrieved September 3, 2026). How the field stands on this axis is in the comparison.

09Sending

The minutes go to the people who were there

The way from the end of the meeting to the recipients

Three stations: first, four separate audio tracks end at one line, the meeting is over. Second, a draft of the minutes with four blocks is ready. Third, four pigeonholes carry the participants' names, and the same sheet lies in each; on one the name tag is still open because the address is missing.

  1. 01The meeting ends

  2. 02The draft is ready

  3. 03The pigeonholes carry the names

The open tag is the address: where the meeting only had a name, Raven looks it up in your address book. If it finds none, you enter it once and it is there the next time. The mail goes out through Raven's mail route, not from your own mailbox.

The usual way after a meeting is: export the minutes, switch to the mail program, piece the distribution list together from the calendar entry, hope nobody is missing.

In Raven you open the send dialog from the meeting. The participants are already in the list, because they were in the call. You tick who should receive it, decide whether the transcript and the link to the recording go along, and send it. Corrections to the text you make beforehand, in the minutes themselves.

The message leaves Raven through the platform's mail route. So it recognizably comes from Raven and not from your own mailbox, and Raven asks for no permission to your mailbox for it. We consider that the cleaner design: a tool that may write from your mailbox in your name can do so from the day of the connection onwards.

The minutes are attached to the message as a file, as a PDF or as Markdown.

10Try it

The next minutes write themselves from the conversation

You create an account, get a room and send the link. Your guests open it in the browser, enter a name and are in, without an account and without installing anything. After the call, the recording and the transcript with names are ready, and you start the draft of the minutes right from there.

Start free

The app is in German for now.

See the measured error rates
11Questions

Questions about AI meeting minutes

How can I have AI write meeting minutes?

You need a recording of the meeting. Either the meeting takes place in a tool that records, or you upload an existing recording. From it comes a transcript with speakers, and from that a draft with topics, decisions and action items. Tell the participants beforehand that the meeting is being recorded, and read the draft before you send it.

Which app writes meeting minutes?

There are two kinds. One joins your meeting as an extra participant and records along, for example Fireflies or Otter. The other is the meeting itself and records what already runs through it, as Raven does. The difference shows in speaker separation: a tool that receives a finished mixed recording has to work out the voices from it, a tool that runs the conference itself has one audio track per participant.

How can I summarize meetings with AI?

From the transcript, not from memory. A language model reads the transcribed text and structures it by topics, decisions, action items with the person responsible and the deadline, and open points. What matters is that the transcript carries names and not Speaker 1 and Speaker 2, otherwise no action item can be assigned.

What is the best AI for summaries?

For meeting minutes it is not the language model that decides but what it is given to read. A very good model on a poor German transcript produces well-worded minutes that are wrong. So ask first for the measured error rate on German and for where the speaker names come from, and only then for the model.

How can I get AI meeting minutes for free?

Almost every provider in the field has a free plan, and each limits it differently: by minutes, by the number of meetings, by length or by storage. At Raven the free plan covers five meetings a month with up to four people and up to 45 minutes each, with recording, speaker attribution and summary, but without video recording.

How can I take minutes of a conversation with AI?

Settle consent first, then the technology. Say at the start that the conversation is being recorded and what for, and ask whether anyone objects. Then you record, have it transcribed and generate the draft. At the end you read it through: are the responsibilities right, are the deadlines right, has an open point stayed open.