Why transcription runs on your Mac

There was a period where running speech recognition locally meant accepting worse text. That period is over. The models that fit in a few hundred megabytes now produce transcripts close enough to the hosted ones that the difference stops mattering for a voice note, and they run on the neural engine in a Mac you already own.

So the decision to send audio somewhere is no longer a technical requirement. It is a product decision, and it has costs.

What sending audio costs you

Every hosted transcription service charges by the minute. That cost has to land somewhere, which is why those products have usage limits, or tiers, or a monthly minute allowance you have to think about. A local model has no per-minute cost, so there is nothing to meter and nothing to ration.

It also means your audio exists on someone else’s disk, at least briefly. Most vendors have a retention policy, and most retention policies are honest. That is still a larger surface than not sending it at all.

And it means the product stops working on a plane.

What running locally costs you

Honesty requires the other column.

You download a model. Mochi’s smallest is 227 MB and understands English. The standard one is 336 MB and understands 25 languages. The full one is 483 MB and holds up better in a noisy room. Meeting mode needs one more download, 14 MB, once. Those are the packaged sizes, not rounded figures.

You spend some battery. Recognition on the neural engine is efficient, but it is not free, and a laptop doing continuous recognition will notice.

You get the model that shipped, until you update. There is no silent improvement behind an API.

Where the line lands

Mochi runs recognition and speaker separation on your Mac, and it sends nothing to do it. The tidied version of a note is written by a model you pick and pay for directly, using your own key, going straight from your Mac to that provider.

That leaves exactly two things touching a first-party server: your account record, and the photo you upload once to make your pet, which is deleted right after. Both are listed on the privacy page rather than buried in a policy.