Private transcription on a Mac: local models for audio that should not leave the room
Some conversations should not be sent anywhere: a medical appointment, a legal discussion, a board meeting before the announcement. For those, the useful question is not how good the cloud model is but whether the audio can stay on the machine at all. On an Apple silicon Mac, it can.
What runs locally
On an Apple silicon Mac the app can download two models and run them on the machine. Whisper Large v3 Turbo does the transcription. Qwen 3 8B does the answers. They are downloaded once and then work on the Mac itself, for sessions whose audio should not leave the room.
Choosing between cloud processing and the local models is a per-workflow decision. A weekly standup does not need the same treatment as a conversation with a lawyer, and nothing forces you to pick one for everything.
What goes where, in each mode
It is worth being exact, because “private” is used loosely.
- Cloud session: microphone audio — and on the Mac, optionally system audio and a captured screen — is streamed to OpenAI's Realtime API to produce transcripts and answers. Attachments are sent as context.
- Local session on the Mac: the downloaded models process the audio and text on the machine.
- Either way, transcripts, answers, sessions, tasks and notes are stored on the device. There is no cloud sync.
- The app's backend keeps your account, purchases, credit balance and usage records. It does not receive or store your audio, transcripts, answers or attachments.
The trade-offs, honestly
A model that fits on a laptop is smaller than one that runs in a data centre. Answers from an 8-billion-parameter model are good for a lot of things — pulling out what was decided, explaining a term, drafting a follow-up — and noticeably less capable than a large cloud model on anything that needs broad knowledge or long reasoning.
The downloads are measured in gigabytes, and the models use real memory and processing power while they run. A recent Apple silicon Mac handles it; an older one will be slower. This is also why it is a Mac feature: phones do not have the room.
On iPhone and iPad
The large local models are Mac only. On iPhone and iPad, transcription can still come from Apple's speech recognition running on the device rather than from OpenAI's model. The answers, in that case, still come from the cloud model — so the words you speak are transcribed on the phone, and the help is not.
Private is also about the people in the room
Keeping audio on the machine protects it from third parties. It does not change the fact that you are capturing what other people say. They still deserve to know, and in many places the law says so. The guide on recording and transcribing other people covers what to tell them.
Which Macs can run the local models?
Apple silicon Macs. The models are downloaded once and run on the machine; newer Macs run them faster.
Can the local models run on iPhone?
No. On iPhone and iPad, transcription can use Apple's on-device speech recognition, but the local answer model is Mac only.
Does the app's server ever see my transcripts?
No. The backend handles account, purchases and credits. Transcripts and answers are stored on the device.
