I wanted notes that write themselves from the call, without the call audio leaving my Mac. MeetingNotes sits in the menu bar. When a calendar event is in progress and a call app is holding the microphone, it starts recording. When the call app lets go of the mic, it waits 45 seconds and stops.
How it works
Is there a call? Core Audio will tell you which processes are using the microphone. MeetingNotes matches them by bundle ID, helper processes included, against Zoom, Teams, Meet, Slack, FaceTime and Webex. That turned out to be a better signal than the calendar. Test events with no invitees and no video link used to get skipped.
Who said what. There’s no speaker detection. The mic is “Me”, and system audio, captured with a Core Audio process tap, is “Them”. Each stream gets its own transcriber running Apple’s on-device SpeechAnalyzer.
The note. Afterwards, my rough notes and the transcript go to a model and come back as Summary, Notes, Decisions, Action Items and a collapsed Transcript. Claude is the default. Apple Foundation Models, Ollama, OpenAI or any compatible endpoint also work. The note lands in Obsidian with attendees as wikilinks and action items as checkboxes with due dates, in the format the Tasks plugin reads. A summary takes about ten seconds. The audio is deleted once the note is written.
What fought back
Silence. The obvious way to keep the speakers out of the mic is Apple’s voice-processing echo
cancellation. On macOS 26 it handed the transcriber a silent channel, and stopped the process tap
from delivering anything at all. I dropped it, downmixed to mono, and handled echo myself. While
recording, the app checks that both streams are alive and restarts the tap once if one goes quiet.
A --diagnose mode runs the real capture code against a file made with say.
Echo. Without Apple’s version, laptop speakers leak the other side of the call into the mic and “Me” starts saying everyone else’s lines. On a 33-minute stand-up, about 320 of 560 “Me” lines were echo. Matching words against the “Them” transcript helped a bit. Loudness worked better. In 50 ms frames, the detector learns how much system audio normally bleeds into the mic, and keeps a “Me” line only when the mic was clearly louder than that. Doubtful lines are held out of the live transcript and judged again against the whole recording at the end. On the same stand-up that left 138 “Me” lines instead of 233, and kept 28 of the 31 things I actually said.
Evening meetings filed under tomorrow. ISO 8601 formatting defaults to UTC, and a 7:31 PM Central call is 1:31 AM the next day in UTC.
The last sentence. The notepad saved after a one-second pause, so whatever I typed just before stopping was lost. It now updates in memory on every keystroke and only batches the disk write.
Odds and ends
It’s signed with a self-signed certificate. That sounds pointless until macOS forgets the microphone and screen-audio permissions on every rebuild of an unsigned app. A packaging script zips it up for my other Mac.