The dictation apps that feel instant, and the ones that do not.
Nobody abandons dictation because of a benchmark. They abandon it because of a pause that makes them doubt the sentence they just said.
The threshold, not the number
Speed in dictation is not a score you want to be high on. It is a line you are on one side of. Below roughly a quarter of a second, the words feel like they came from you and you carry on to the next sentence without thinking about it. Above it, you wait to see what appears.
That waiting is where dictation dies, and it dies quietly. You start speaking more carefully so you will not have to fix things. You pause between clauses to let it catch up. You say shorter sentences. Every one of those adjustments makes you slower than typing, and none of them feels like a decision to stop using the tool.
Where the time goes
A cloud dictation app buffers your audio, sends it, waits for a machine somewhere to answer, and pastes the reply. Most of that is not computation, it is travel and queueing, and it has a floor that your hardware cannot lower. It is also why the same product feels fine at a desk and unusable on a train.
Running the model on your own silicon removes the trip entirely. Zumbo composes a sentence in about a fifth of a second on the Mac itself, with no request leaving the machine, which is why it behaves identically on a plane, in a basement and on a bad connection.
What it feels like at that speed
Hold it to talk, tap it to toggle, Escape cancels. Pick the key in Settings.
You stop treating dictation as a mode you enter. It becomes the way you put a sentence somewhere, no different from typing one, and you use it for a six word reply rather than saving it for a paragraph. That is the actual test of whether a dictation app works: not how impressive the demonstration is, but whether you reach for it without deciding to.
The second clock nobody measures
Latency is only half the cost. The other half is every wrong word you have to go back and fix. Three of those in a paragraph and you have spent more time than you saved, no matter how fast the text arrived.
That is why speed and vocabulary are the same conversation. Twenty one packs cover the language of a field so the common terms are right from the start, and a word corrected once with Teach is heard your way from then on. The fastest dictation is the one you do not have to correct.
Why there is no comparison table on this page
Because none of the other apps publish a latency figure that can be opened and checked, and putting a number next to a competitor's name without a source would be the exact thing the rest of this site refuses to do. The number here is ours, for our own app, measured on device.
Related: speaking against typing speed, with the studies, dictation with the Wi-Fi off, and Zumbo against Wispr Flow.
What does 200 ms actually mean here?
The time between finishing a sentence and that sentence appearing where your cursor is, composed on your Mac. It is not a benchmark of anybody else's product.
Why does the threshold matter more than the number?
Because below it you keep talking and above it you start waiting. Once you are waiting, you dictate carefully, and careful dictation is slower than typing. The failure is behavioural, not technical.
Does a faster Mac make it faster?
Somewhat, but the architecture decides most of it. A round trip to a server has a floor set by the network, and no amount of local hardware helps.
Does it stay fast offline?
It is the same either way, because there is no network in the path. The engine ships inside the app and runs on your Mac.
Is accuracy traded for speed?
They are separate problems. Speed comes from running locally; accuracy on your words comes from the vocabulary packs and from Teach.
Free for 3 days, everything unlocked. Then $15 once. macOS 14 or later, Apple silicon.