How to transcribe meeting audio on your Mac.

1. Capture both sides. A meeting is your microphone plus the other people’s voices, and on a Mac those voices arrive as system audio. Record only the microphone and you have half a conversation. How macOS lets an app take both is the mechanical half of this.
2. Turn speech into text. Speech recognition produces the words and the timings against them. It cannot recover what part one failed to capture, which is why the order matters.
3. Make it a document. Timestamps, who said what, and a summary that still makes sense in a week. This is the step most do-it-yourself pipelines skip, and it is usually the thing you actually wanted.

Uploading the file to a service and getting a transcript back does the same job, often very well, and does it somewhere else. Which one you want is a real decision, and it is its own page.

• Apple silicon, realistically. Transcription models lean on the Neural Engine and the GPU. An Intel Mac will run them and will take considerably longer.
• Less disk than you expect. An hour of compressed meeting audio is tens of megabytes. A local speech model is a one-time download, a few hundred megabytes to a few gigabytes depending on how large a model you want.
• Permission to record system audio. macOS gates it, and the prompt asks for Screen Recording even when nothing on screen is being recorded. That surprises people, and there is a reason for it.
• A decision about consent. No framework and no service asks the room on your behalf. Tell people you are recording.

• The recording caps everything after it. Speakers instead of headphones, a laptop microphone across a table, two people talking at once. No model recovers what was never captured cleanly, and this is the single biggest lever most people never touch.
• The model is the size your machine can hold. A datacenter runs a larger one on hardware that is not on your desk. On clear speech the difference is small. On strong accents, crosstalk and specialist vocabulary it widens. One measured run against three hosted services, with the method written out, is on the benchmark page.
• Speaker labels are partly a guess. Separating your voice from the far end is easy, because the microphone and the system audio arrive as two separate streams. Telling two people apart inside one of those streams is inference, and a calendar invite is what turns speaker 2 into a name.
• The summary is a second machine. Reading a transcript and writing a summary is a language model’s job, not a speech model’s. Either it runs on your Mac, slower and smaller, or the text goes to one that does not. Most tools that describe themselves as local do the second thing, so it is worth reading the fine print.

Aside is route three. It records and transcribes on your Mac, and the speech model ships inside the download, so the first meeting works with no account, no upload and nothing metered. Around 99 languages, detected automatically, and it works with the wifi off.

Writing the note is the separate step described above. By default that transcript text goes to Claude Haiku through Aside cloud. Point it at your own API key and it goes to your account instead, or at Ollama on your own machine, in which case nothing goes out at all. The full list, including the uncomfortable line, is under privacy on the front page.

Your meeting notes are files on your Mac, which is the other half of the argument: a transcript you cannot get out of an app is a transcript you are renting.

Last updated 13 August 2026.