Why your dictation app should never need the cloud

Here is the thing about dictation that makes it different from almost every other AI product: the input is not a prompt you crafted for a machine. It is the sentence you were about to write to a colleague, a client, a friend. Your unsent email. Your half-formed opinion of the meeting you just left. Dictation software sits closer to your inner monologue than any other tool on your Mac.

That's why we think the question "where is this processed?" matters more here than anywhere else.

What actually leaves your machine

When a dictation app processes speech in the cloud, three things travel: your voice itself (a biometric), the words you said (the content), and the context around them (which app, what time, how often). Every one of those is covered by a privacy policy, and privacy policies are real commitments made by mostly decent companies.

But a policy is a promise, and promises have failure modes: the acquisition, the breach, the subpoena, the "we've updated our terms" email, the retention setting someone forgot. None of these require malice. They just require time.

The architectural answer

There is exactly one privacy guarantee that survives all of those failure modes: the data never existed anywhere else. If your voice is transcribed on your Mac, rewritten on your Mac, and typed into the frontmost app by your Mac, then there is nothing to breach, nothing to retain, nothing to hand over. The guarantee isn't in a document. It's in the packet capture, where you can check it yourself: run Murmur with the network unplugged and nothing changes.

This used to be a trade-off. On-device meant worse. It doesn't anymore.

The hardware already caught up

Apple silicon ships with a neural engine that idles through most of its life. Apple's on-device speech recognition is fast and accurate, and Apple Intelligence can rewrite a rambling sentence into a professional one in well under a second, locally. The models are on your disk, maintained by the same company that maintains your operating system, improving with every macOS release.

That's the quiet unlock: the cloud stopped being necessary for this job. What remains is habit, and business models that need a server in the loop to justify a subscription.

The economics follow the architecture

Notice how the pricing lines up. Cloud dictation apps charge monthly because they have inference bills to pay: your subscription is largely paying for GPUs. On-device apps have no marginal cost per word, which is why Murmur is free and even the paid local apps like Superwhisper can offer lifetime licenses. When someone charges you monthly for dictation, part of what you're renting is their decision not to use the computer you already own.

What to demand from any dictation app

Whatever you end up using (our comparisons are honest about the alternatives), we'd suggest three requirements:

• It should work offline. Not as a degraded mode. Identically. That's the proof the pipeline is local.
• It should not need an account. An account is a database row about you. Transcribing your own voice on your own computer requires none.
• It should refuse your secrets. Password fields exist. A dictation app should detect secure input and decline to type there, the way Murmur does.

Your voice is the rough draft of your thinking. Software that handles it should treat that as the sensitive material it is, and the only architecture that guarantees it is the one where your words never leave home.