Your voice never needed the cloud
There’s a moment in every voice app’s privacy policy where the prose gets careful. You can feel the lawyers leaning in. Audio may be processed on our servers to improve service quality. Transcripts may be retained to enhance our models. We take your privacy seriously.
Here’s what those sentences work hard not to say outright: when you dictate to the cloud, your voice leaves your machine. The rough draft of the email you never sent. The message you rewrote three times because the first version was too honest. The idea you talked out at midnight and deleted by morning. All of it streamed off your computer at the exact moment it was least finished and most yours.
For about ten years, this was a fair trade — because there was no other option. It’s worth understanding why that was true. And why it isn’t anymore.
Why dictation moved to the cloud in the first place
Good speech recognition used to be heavy. Handling real speech — accents, background noise, people talking the way people actually talk — took more memory and power than a laptop could spare. So the industry settled on a design. Your microphone became a sensor. The intelligence lived in a data center.
That design had real strengths, and honesty means listing them. Server models could be huge. Their makers could update them overnight. They could learn from millions of users at once. Cloud dictation got good, and it got good fast, because everyone’s audio fed the same improvement loop.
But the design had a cost that never went away, because it couldn’t. The cloud is other people’s computers. Every privacy promise a cloud voice product makes is a policy — retention limits, encryption, “we don’t train on your data.” And policies change. Companies get bought. Terms get updated. A subpoena arrives. An employee leaves a storage bucket open. Nobody has to be a villain. It just takes time, incentives, and the audio being there.
There’s a useful rule in security: data that isn’t collected can’t leak. The strongest privacy guarantee isn’t a promise about what happens to your voice on a server. It’s your voice never reaching a server at all.
What changed: the data center in your laptop
Two things happened at about the same time, and few people noticed either one.
First, the models got far more efficient. Researchers spent years squeezing what speech recognition knows into smaller, faster networks. The accuracy that once needed a server rack now fits in a download of a few hundred megabytes.
Second, the laptop grew a brain built for this exact job. Apple silicon Macs ship with a Neural Engine — a chip made for running machine-learning models. It sits idle most of the time. Run a modern speech model on it and it doesn’t just work. It works fast — faster than real time — while sipping battery.
Put those together and the old bargain expired. On-device recognition is no longer the weak option that misses words and demands robot-speak. On a modern Mac, it’s the right way to do it, full stop. Your audio goes from the microphone to the Neural Engine to text at your cursor, and the story ends there. No network round-trip. It works on a plane. It works when your Wi-Fi doesn’t. And a data center that’s zero millimeters away is hard to beat for speed.
The questions worth asking any voice app
If you’re weighing a dictation tool — ours or anyone’s — the marketing page will say “private” no matter what. These questions separate real design from adjectives:
- Where does transcription run? Not “is it encrypted” — encrypted on its way to whom? If the answer involves a server, everything after that is policy, not physics.
- Does it work with the network unplugged? This is the nice thing about on-device claims: you can test them. Turn off Wi-Fi. Dictate. A cloud product fails this test in one sentence.
- What happens to audio after transcription? The best answer is that there’s nothing to answer. Your machine handled the audio in memory, and it was never anyone’s to keep.
- Is there telemetry, and what’s in it? “We don’t send your audio” still leaves room for sending everything about your audio. Ask what leaves the machine. Period.
- If the app offers AI cleanup, who runs that model — and who decides? This one has a fair range of answers. Cleanup with a local model keeps the whole pipeline on your machine. Some people want a big cloud model for heavy rewriting. That’s fine — if the user chooses it each time, and it’s not the product’s silent default.
That last line matters more than any single feature. Privacy isn’t refusing you the cloud. It’s the cloud being a door you open — with your eyes open, when you want what’s behind it. Never a pipe hooked up before you asked.
Where we stand
We built Clabrate on the unfashionable version of this. The private path isn’t our premium tier or a compliance checkbox. It’s the default and the point. Speech recognition runs on your Mac’s Neural Engine, start to finish. The model downloads once, then works — network or no network. Text cleanup runs on a local model. No telemetry. If you want to bring your own cloud provider for heavier rewriting, you can. Your key, your choice, made in the open. And if you never do, nothing you say ever needs to leave the room.
We won’t pretend this is the easy design. It’s harder to build. It can’t quietly learn from your audio to improve itself. We don’t get a dashboard of your usage to admire. But the other path meant asking you to stream your least-finished thoughts to a data center — and trust the paragraph where the prose gets careful.
Your voice never needed the cloud. It just needed the hardware to catch up. It has.