How to Transcribe Zoom, Teams, and Meet Locally

Why live transcription

Taking notes in a meeting is a trade-off: either you miss what is said or you write a pile of text you never use. Live transcription is simple: words appear as people talk, you listen, and you clean up afterwards.

Audio Note has three live modes:

Mode	What it captures	Good for
Microphone realtime	Built-in or external microphone	In-person meetings, dictating, classes
App realtime	Audio from a chosen app	Zoom, Teams, Meet, and other online meetings
Global realtime (beta)	Everything the system plays, shown in a caption window	Switching between apps, livestreams, videos

All three work on Mac and Windows. Live transcription starts at the Standard plan; see the pricing page.

Local capture vs a meeting bot

A cloud notetaker joins the call as a participant and sends audio to its servers. Local capture does not join anything: Audio Note transcribes what your computer is already playing. Nobody sees an extra attendee, and there is no waiting room to admit.

Transcription runs on your machine with Whisper (whisper.cpp) or Sherpa-ONNX realtime models, and results are stored locally. Recordings are not sent to a transcription server.

Five minutes of setup

1. Permissions

Grant these once, before a real meeting:

Microphone: needed for microphone realtime. Windows does not let apps request it directly; enable it in Windows Settings, then come back and click "Recheck".
Screen Recording (macOS): needed for app realtime to capture an app's audio and window. Enable Audio Note under System Settings → Privacy & Security → Screen Recording.

2. Pick a model

Situation	Start with	Why
Ordinary laptop, long meetings	Realtime model	Low latency, no discrete GPU needed
Strong GPU, want more accuracy	Whisper	With enough GPU, Whisper keeps up in real time
Not sure	Run each for 3 to 5 minutes	Compare latency and mistakes on real audio

On Standard, realtime models cover Chinese and English; Pro adds more languages. Details in the models overview.

3. Run a short test

Play a recorded meeting or start a test call and run two or three minutes. Permissions, audio routing, fullscreen windows, and Bluetooth headset switching are the usual failure points, and finding them before the meeting is much cheaper.

Capturing an online meeting

Zoom, Microsoft Teams, Google Meet, Lark, and Tencent Meeting all follow the same steps:

Start the meeting in its app or browser.
In Audio Note, choose "App Realtime Transcription" and pick the meeting window.
Keep the meeting window on the same screen as Audio Note and out of fullscreen. Apps in fullscreen or on another virtual desktop may not be captured.
Choose the model and language, then start.
Do not mute the meeting on your side. Local capture records what is played back, not the other side's upstream audio.

One caveat: because of system limitations, choosing a single app may still capture the whole system's audio. Mute music, videos, and notification sounds during the meeting so they do not end up in the transcript.

If the meeting runs in a browser and you also need other apps, try global realtime (beta). It captures at the system level and shows a separate caption window. Run it for five to ten minutes first before relying on it.

Should you also capture the microphone?

For a lecture or livestream you only listen to, a single source is cleaner. Add the microphone when your own voice should be in the record.

With speakers instead of headphones, the microphone picks up the speaker output again and you get duplicates. Wear headphones, or enable "Reduce Speaker Echo" (experimental) under Settings → Recording & Save.

During the meeting

Text scrolls automatically; you can pause to look back.
Do not switch models, parameters, or audio devices mid-meeting; that hurts more than a few wrong words.
Treat live text as a draft. Fix names, numbers, and dates afterwards.

After the meeting

When you stop, the transcript is saved. Then:

Fix the obvious errors, such as names and numbers, in the note view.
Open the AI assistant, ask for the key decisions first, then reshape them into owner, due date, and risk.
Export Markdown for people who missed the call, or export subtitles for the archive.

AI processing uses the backend you configured. Local models such as Ollama or LM Studio keep text on your computer; cloud APIs send it to that service. See AI summaries and action items.

Getting better accuracy

Audio quality matters more than the model. For in-person meetings, use an external microphone close to the main speaker, away from air conditioning, keyboards, and speakers.

Set the language. Chinese meetings in Chinese, English in English. For mixed speech, test which setting works better on your own sample.

Add speakers afterwards. Audio Note does not separate speakers automatically yet; that feature is still in development. If you need to know who said what, type speaker names at the start of paragraphs in the editor after the meeting.

Split long meetings. There is no fixed time limit, but very long transcripts make editing, search, and AI processing heavier. Save important meetings by topic.

Troubleshooting

Symptom	Check first
Captions stay empty	macOS Screen Recording or Windows microphone permission
Meeting window not listed	Is it fullscreen or on another virtual desktop?
Music and alerts in the transcript	Mute other apps
Your words appear twice	Speaker echo; use headphones or enable "Reduce Speaker Echo"
Text falls behind	Model too heavy for the device; switch to a realtime model or smaller Whisper

FAQ

Will others know I am transcribing? No bot or participant appears in the meeting. Whether you must tell people you are recording depends on local law and company policy; check yours.

Does it work offline? After the model is downloaded, transcription does not need the internet. Sign-in and cloud AI features may still go online.

How long can a meeting be? No hard limit; it depends on your computer. Save long meetings in parts.

Is live transcription on the free plan? No. It starts at Standard. You can try file transcription on the free plan first.