Meeting Transcription: How It Works and What to Use

Meeting transcripts come from four sources: the meeting platform's own transcript, a note-taker bot that joins the call as a participant, an app that records the audio on your own computer without joining, and a recorder placed in the room. The platform transcript is the simplest choice when everyone is on one service and your plan includes it. The other three cover calls and in-person meetings, but each one trades off cost, consent and accuracy. Whichever you pick, tell everyone the meeting is being transcribed. Some places require the consent of every participant.

Most meeting transcripts now come from the meeting platform itself, usually on paid plans

The platform is the default source because it already has the audio, the speaker names and the meeting record. Microsoft Teams tells all participants when live transcription starts ¹. Google Meet shows attendees a notice before they join, when transcripts are active ². Google also says that transcription is on by default, but someone in the meeting has to start it ³, so a meeting can run without a transcript if nobody starts one.

Access depends on the edition. Google's administrator help lists the Google Workspace editions that support meeting transcription: Business Standard, Business Plus, Enterprise Standard and Enterprise Plus, and Education Plus ⁴. Check your edition in the Admin console before you plan around the feature.

Teams has a platform limit as well. Live transcription is only available on the desktop app ⁵. On the mobile app you can open a transcript only after the meeting ends ⁶. If you join from a phone, expect to read the transcript later.

Zoom makes audio transcripts from cloud recordings. A university's Zoom guide places the option under Advanced Cloud Recording settings ⁷, so cloud recording has to be set up first. We could not open Zoom's own plan page to confirm the current plan list, so check the requirement in your Zoom account before you rely on it.

Platform transcripts are the best default when the whole meeting runs on one service and your edition includes the feature. They are not the answer when people join from several services, call in by phone, or meet in a room.

Note-taker bots, botless desktop apps and room recorders capture the audio in different ways

A note-taker bot joins the call as a participant and listens through the meeting service. Everyone in the meeting can see it. The bot's transcript is generated on its service, so read that service's privacy terms before you send it a confidential call.

A botless app captures the audio your computer plays and records through its microphone. Nobody sees an extra participant. The trade-off is that the app listens to one device, and it must be running on that device when the meeting starts. For the short answer on whether a meeting can be transcribed with no bot at all, see can you transcribe a meeting without a bot joining.

Notion's AI Meeting Notes is an example of the botless design. Its browser version can only pick up audio from your microphone, so Notion describes it as best for in-person conversations ⁸. The feature requires a Business or Enterprise plan ⁹.

Platform support varies. Superwhisper lists meeting transcription as Mac only on its home page, while its dictation is listed for Mac, Windows, iOS and Android ¹⁰. If you run Windows, check each app's current platform list before you choose a botless tool.

A room recorder is a device or phone on the table that records everyone present. It needs no account on the call, but the speakers are at a distance from the microphone. The section on crosstalk and microphones below shows what that distance does to accuracy.

A transcript is the record of what was said; an AI summary is an interpretation of it

A transcript is the words, in order, from the audio. A summary is a model's rewrite of the transcript, with decisions and action items pulled out. Each step can go wrong in a different way, so you should check both against the audio before a decision depends on them.

The clearest evidence comes from clinical notes, where AI scribes produce summaries of consultations. In a 2026 UC Davis pilot of AI scribe notes, 11.5 percent of the 356 reviewed notes contained hallucinations, meaning content the recording did not support ¹¹. A meeting summary is not a clinical note, but the same risk applies: a summary can drop a point, or add one the speakers never made.

The transcript itself can invent text. In a 2024 study, roughly 1 percent of Whisper audio transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ¹². The authors tested Whisper as it stood in 2023, so that figure describes that version ¹³. A separate 2025 study found that when non-speech sounds were played to Whisper, a set of hallucinated phrases appeared frequently ¹⁴. Long silences, hold music and room noise are where to look first.

For a record that matters, keep the audio, read the transcript against it at the points where a decision was made, and treat the summary as a draft. The guide to how to turn a meeting transcript into minutes covers that step.

Decide the style before the meeting starts, because it changes what the transcript keeps. A verbatim transcription keeps fillers, false starts and repeated words, which a legal or research record may need. A clean verbatim transcript removes them, which reads better as a record of decisions. Our guide to verbatim and clean verbatim transcription sets out when each one fits. A platform or bot may not let you choose, so check the output before you rely on it.

Crosstalk, laptop microphones and compressed call audio are what lower meeting accuracy

Accuracy in a meeting depends on three things: how far the microphone is from each speaker, how many people talk at once, and how the call was compressed. Accuracy is measured as word error rate, where lower is better. The same model can score very differently across these conditions.

Distance matters most. On the AMI meeting corpus, Whisper Large V2 scored 16.9 percent word error rate when each speaker had a headset microphone ¹⁵. With a single distant microphone it scored 36.4 percent ¹⁶. That makes the distant recording about 2.15 times the headset error rate for the same model ¹⁷. A laptop on the table, far from half the room, is closer to the distant case than the headset case. Apple gives the same advice for dictation: a headset microphone may help in a noisy room or one with excessive echo ¹⁸.

Overlapping talkers are the second cause. On the CHiME6 conversational test set, Whisper Large V2 had a word error rate of 25.5 percent ¹⁹, against 2.7 percent on clean read speech from LibriSpeech ²⁰. The test sets are different, so the gap is a comparison of conditions, not a single controlled test. Still, it shows why a meeting where two people talk at once is harder than a read paragraph.

Compressed call audio is the third cause. The Bluetooth Headset Profile carries voice at 8 kHz with the CVSD codec or at 16 kHz with the mSBC wideband codec ²¹. The 8 kHz option is the narrower band, and it keeps less of the sound the recognizer needs.

Read leaderboard figures with care. The Open ASR Leaderboard removes punctuation, casing and disfluencies such as fillers before it computes word error rate for its normalized transcripts ²². A meeting transcript that keeps every "um" and restart will score worse than those tables. Treat leaderboard scores as best-case accuracy on cleaned-up text.

Three habits raise accuracy more than any setting you can change. Use a headset when you can, because distance made the biggest difference in the same-model comparison above. Ask people to speak one at a time, because overlapping talk is the other major cause. And record a two-minute test before a long meeting, then listen for dropouts and muffled speech, so you know the transcript will be usable before it matters.

Everyone in the meeting should know it is being transcribed, and some laws require consent

Consent often decides the method. Platform transcription notifies people by design. Teams notifies all participants when live transcription starts ¹, and Google tells attendees before they join when transcripts are active ². A bot or an app may not do this for you. Announce the recording at the start of the meeting, and say it again if someone joins late. A short line covers most cases: say that the meeting is being transcribed, say where the transcript will be stored and who will see it, and ask anyone who objects to say so before the call goes on. If someone objects, stop the transcription, and do not start it again for that meeting without asking.

The law differs by place. In the United States, the federal Wiretap Act allows recording when a party to the conversation consents, and that is the one-party rule ²³. California's Penal Code 632 is different: it punishes recording a confidential communication without the consent of all parties ²⁴. If your meeting includes people in California, get everyone's consent. The terms are defined in the all-party consent and one-party consent entries, and our guide to recording consent laws covers the rules by state.

Speaker labels raise a separate question. Under the GDPR, Article 9 prohibits processing special categories of personal data, and biometric data used for the purpose of uniquely identifying a person is one of them ²⁵. Whether a voice model used for speaker labels is biometric data depends on whether it identifies people. This is a question for your privacy lawyer, not a settled rule. Our guide to voice biometric data explains the terms.

Local and open-source meeting transcription keeps the audio on your own machine

A local tool keeps the audio on your own computer, so nothing goes to a meeting service or a cloud API. That suits confidential or regulated meetings, but it does not remove the consent duties above.

Whisper.cpp is a C/C++ port of the model. Its README describes Apple Silicon as a first-class target ²⁶, so Apple Silicon Macs are a practical place to run it. Faster-whisper is a reimplementation of Whisper on the CTranslate2 engine, and it is released under the MIT License ²⁷. Both run without sending audio to a company, which is the point of a local tool.

Speaker labels need an extra step, covered in more depth in our guide to timestamps and speaker labels. WhisperX adds word-level timestamps and speaker diarization on top of Whisper-based recognition ²⁸. Speaker diarization is the task of working out who spoke when. A 2019 Google study on physician-patient conversations cut word-level diarization error rate from 15.8 percent to 2.2 percent with a joint model, against a conventional baseline ²⁹.

Hosted services differ in what they label. Google's documentation says speaker diarization is not available for every language and tells you to check its language support page ³⁰. Check the labels you need, and the languages you speak, before you choose a service.

For a full walkthrough of local setup, read our guide to open-source speech recognition and on-device speech recognition.

Pick the tool by where your meetings happen and who needs the transcript

Start with the place. If the meeting runs on one platform, use its transcript. If it runs on several services, or on phone calls, use a botless app or a bot, after the consent step. If it is in a room, use a recorder close to the speakers, or record the room and transcribe the file later.

Who needs the transcript decides how much checking it needs. A personal reminder can be a rough draft. A record that others rely on needs the audio checked at each decision, and a person who approves the final text.

To work through the choice for one meeting, use this order:

1. List who will attend and which services or rooms they will use.
2. Confirm that your edition or plan includes transcription on that platform, and that an admin has enabled it.
3. Pick the method from the table near the top of this guide.
4. Announce the transcription at the start, and ask for objections before you begin.
5. Record a short test and listen to it, checking for dropouts, overlapping speech and muffled voices.
6. After the meeting, check the transcript against the audio at each decision, and keep the audio if the record matters.

Three more pages go deeper on the choices above. The guide to Zoom, Teams and Google Meet transcription walks through the platform settings. For a meeting that already happened in person, see how to transcribe an in-person meeting. If you are weighing a dedicated recorder, read whether AI transcription devices are worth it. If cost decides it, see whether there is a free meeting transcription app.

If you want to dictate your own notes after the meeting, Yap, which we make, is an on-device dictation app for Mac and Windows. For a free way to turn a recording file into text in your browser, use our audio to text tool. For the basics of the whole subject, see our transcription overview, and the guide to how to transcribe audio to text covers the file-based steps.