Transcribing Zoom, Teams & Google Meet

Bot-based tools join the meeting as a participant. They need calendar access, meeting-platform permissions and often an admin to approve the integration for your whole workspace. Everyone sees the bot in the participant list. You get a summary afterwards.

Local capture records the audio your computer is already playing, on your machine. No bot joins, no calendar integration, no admin approval — and because it is not tied to any platform's API, it works with anything that makes sound.

Bots are better for recurring internal meetings where everyone expects one. Local capture is better for everything else, and it is the only option that helps you during the call rather than after it.

Local capture operates below the application layer, so platform support is not really a feature list.

The only real requirement is that the sound reaches your speakers or headphones. If you can hear it, it can be transcribed.

The part that actually matters is capturing two streams at once.

System audio is everyone else on the call — the output from your speakers.

Microphone audio is you.

Both are captured simultaneously and merged into a single running transcript, so the record reflects the conversation rather than one half of it.

This is the piece browser extensions struggle with. Reaching a microphone from a browser is straightforward; capturing the system's output device generally is not, which is why those tools usually ask you to share a tab and tick "share audio" — and why they hear nothing when the call is in a desktop app.

Accuracy depends far more on your setup than on the model.

Use a headset. Biggest single factor and it is free. Laptop speakers into a laptop microphone creates echo, and the echo gets transcribed too.

Set the right language. A model running in the wrong language mode produces confident nonsense.

Expect jargon to fail. Product names, internal acronyms and unusual surnames are the common errors. Normal, not a fault.

Check your connection. Streaming transcription sends audio continuously; a weak connection causes dropouts and lag.

Cross-talk degrades everything. Meetings with clean turn-taking transcribe dramatically better than chaotic ones.

Scroll back instead of asking someone to repeat a figure.

Answer a question that lands on you, drafted from documents you uploaded beforehand.

Ask about a shared slide by capturing a screenshot and running it through a vision model.

Keep your notes private — the assistant window can exclude itself from screen capture, so it is not projected when you share.

Recording and transcription law varies by country and by state, and several jurisdictions require every participant to consent. A local tool that announces nothing makes it easy to forget, so the responsibility sits with you rather than the software.

The practical version: say at the start that you are taking AI-assisted notes. Four seconds, almost always met with indifference, removes the entire problem. For anything under a confidentiality agreement, check whether sending content to third-party AI providers is permitted before you upload it.