A choice hiding inside a feature
When you turn on dictation, you're making a decision most apps never surface: where the actual work happens. Your voice has to be turned into text somewhere, and there are only two somewheres. Either the conversion runs on the chip in the device in your hand, or your voice is sent over the internet to a company's servers, processed there, and the text sent back. That fork — on-device vs cloud dictation — looks like an implementation detail. It quietly determines what happens to your voice, whether the tool works when you're offline, and how it feels to use.
For years the choice barely existed, because phones and laptops weren't powerful enough to do good speech recognition locally, so everything went to the cloud by necessity. That's changed. The models have gotten small and efficient enough to run on the device, which means the fork is now real, and worth understanding before you pick a side.
What "the cloud" actually does with your voice
In a cloud setup, the audio of you speaking leaves your device. It travels to a server, where it's transcribed, possibly logged, and the result returned. This is not sinister by default — it's just how the plumbing works — but it has consequences that are easy to ignore until they matter.
The first is exposure. Your voice, and everything you dictate, passes through and often rests on infrastructure you don't control. What's retained, for how long, who can access it, and whether it's used to improve the company's models are all governed by a privacy policy you didn't read and can't enforce. For a grocery list this is academic. For a therapy journal, a medical note, a legal matter, a private message, or anything you'd be uncomfortable having sit in a database, it's the whole ballgame. The audio of your actual voice is among the more sensitive things you can hand over, and dictation hands it over constantly.
The second consequence is dependence. Cloud dictation only works when you have a connection. On a plane, in a basement, on a bad-signal train, in a foreign country with data off — the moment you most want to capture a thought hands-free, the tool is dead. And every word makes a round trip to a server and back, which adds a small but real delay between speaking and seeing text, and a hard dependence on someone else's servers staying up.
What "on-device" changes
When transcription runs locally, the audio never leaves. It's captured, converted to text on the same chip, and discarded — the words exist on your device and nowhere else unless you choose to put them somewhere. This isn't a promise backed by a policy; it's a property of the architecture. There's no server log of your voice because there was no server. That's a stronger guarantee than any privacy statement, because it's not a commitment to behave well — it's an inability to do otherwise.
Working locally also means working everywhere. Airplane mode, dead zones, the subway, a cabin with no signal — none of it matters, because nothing needs to be sent anywhere. The thought you have offline is just as catchable as the one you have on wifi. And because there's no round trip, the text tends to appear with less lag; the processing is happening inches from your mouth, not a continent away.
There's a subtler benefit too: independence from a vendor's decisions. A cloud service can change its pricing, alter its terms, deprecate the feature, or simply go down, and your dictation goes with it. A model running on your own device keeps working on its own terms — it can't be metered, rate-limited, or quietly switched off from somewhere else. What runs on your hardware is yours in a way that a service never quite is.
The honest trade-offs
This isn't a free lunch, and it's worth being straight about the costs. Cloud servers are vast; a phone is not. Historically the very largest, most accurate models ran only in data centers, and there were speech tasks — heavy accents, obscure jargon, many languages at once — where the cloud had a real edge. That gap has narrowed dramatically as on-device models have improved, to the point where for everyday dictation in common languages most people won't notice a difference. But if your needs are at the extreme edge of accuracy or you work across many rare languages, it's a real consideration.
There's also a flexibility argument for a hybrid stance: keep everything local by default, but allow an optional connection to a more powerful service for the rare job that needs it — ideally one that goes directly from your device to the provider you chose, not through an opaque middleman. That keeps the privacy and offline benefits as the standard while leaving an escape hatch you control and can leave switched off.
How to actually decide
The decision comes down to what you dictate and where. If most of what you'd speak is even mildly personal — anything you wouldn't post publicly — the privacy argument for on-device is strong and gets stronger the more sensitive your material. If you frequently work without a connection, or you just dislike a tool that stops functioning the moment the signal drops, on-device wins on reliability alone. And if you value the snappiness of text appearing the instant you stop talking, local processing tends to feel better.
The cloud's remaining advantages — the last few points of accuracy on the hardest inputs — matter to a real but narrow set of users. For most people doing most kinds of writing, the calculus now favors keeping things on the device, which wasn't true a few years ago and is worth re-examining if you formed your opinion back then.
The default that should have been there all along
The deeper point is that on-device should arguably be the default, with the cloud as the explicit exception you opt into for a specific reason — not the other way around. For most of dictation's history that wasn't possible. Now it is, and a tool that puts your voice on the device by default is simply respecting a boundary that older technology forced everyone to cross.
Quill is built on the local-first side of this fork. Transcription and cleanup run entirely on your iPhone or Mac, so your voice and your words never leave the device, and dictation keeps working with no signal at all. If you ever want the extra reach of a cloud service for a specific job, you can bring your own key and call it directly — off by default, under your control, never routed through us. If where your voice goes is something you'd rather decide on purpose, you can see how it works at quill.lumenlabs.works.