Offline dictation on a Mac: what actually runs locally

Written 15 August 2026.

"Private" and "on-device" are marketing words. Here is what they have to mean technically, and how to check a claim rather than trust it.

Two stages, and they can live in different places

Modern dictation is two steps, and a tool can run them in different places without saying so:

1. Recognition — audio to text.
2. Cleanup — text to text you would send: filler removed, punctuation added, false starts dropped.

A tool can run recognition on your Mac and still send the transcript to a server for the second step. That is a real architecture and it is not dishonest by itself — but "on-device speech recognition" is a claim about step one only, and it is worth reading it as narrowly as it is written.

The claim worth looking for is per-stage. For Bunso: both stages run on your Mac, out of the box nothing leaves it, and audio is never sent anywhere under any setting. There is one optional exception, off by default — you can supply your own API key and have the cleanup step run at a provider you choose. The privacy page lists every item that can be transmitted and what governs each one.

How to check, in about a minute

You do not have to take anyone's word for this.

• Turn off Wi-Fi and dictate. The bluntest test there is. If it works with the network off, recognition is local. If it fails, it was never local.
• Watch the connections. macOS has no built-in per-app network monitor, but a third-party firewall that prompts on outbound connections will show you what an app talks to and when.
• Check the size of the download. This is the tell people miss. Local speech models are hundreds of megabytes to a few gigabytes. A dictation app that installs in 20 MB and works immediately is doing the work somewhere else — there is nowhere on your disk for a model to be.

What local costs you

Running locally is not free, and a page that only lists the upsides is selling something. The real costs:

• Disk. The models are fetched once, about 2.6 GB together, or roughly 0.5 GB if you decline the cleanup pass.
• A slow first launch. The speech model is compiled for the neural engine the first time it runs — a few minutes — and again after each update. It is expected, it is one-time per version, and any tool doing this honestly has the same cost.
• Hardware requirements. Apple Silicon only, on macOS 15 or later. This is not a licensing choice; the speed depends on the neural engine.
• A quality ceiling that moves slower. A local model is the one on your disk. A cloud service can improve without you doing anything — and can also change, or be discontinued, without you doing anything.

Why offline is worth the trade for some people and not others

If you dictate customer records, medical notes, legal drafts, financial figures, HR matters or anything under an NDA, the calculation is simple: the data cannot be sent, so the tool must not send it. Local is the only shape that satisfies that.

If you dictate on a plane, on bad hotel wifi, on a train through tunnels, or anywhere with intermittent connectivity, local is about working at all — a network-dependent tool fails exactly when you are least able to fix it.

And if none of that describes you, offline is a nice-to-have rather than a requirement, and you should pick on accuracy for the language you actually speak instead. Privacy is table stakes in this category, not a reason to choose one tool over another — which is why this site argues about code-switching and not about privacy.