Since April I've said about 613,000 words to my computer.
That's 8,318 dictations in Wispr Flow between 2026-04-22 and 2026-09-25, a bit under 4,000 words a day. The median dictation is 34 words. And most of them weren't meant for a person at all.
I exported the whole history and counted it, mostly because I wanted to know what I actually sound like. Some of it was useful. Some of it was a little embarrassing.
Most of it goes to an agent
About 73% of those dictations went into a terminal, to a coding agent. Another 11% went to an AI in the browser. The rest is spread across Chrome, Safari, WhatsApp and Notes.
That split makes sense once you see how I work now. I manually don't code a lot nowadays. I validate code, I check the quality, and of course I add my own inputs. Most of my day is directing agents the way you'd direct a team: here's the context, here's the scope, don't touch this part yet, raise the PR and let me know.
That kind of instruction is long, and typing it is where I used to cut corners. For example, 359 of my dictations are over 300 words. The longest ones run past 2,000. I would never have typed those. I'd have typed 40 words and hoped the agent filled in the rest, and then spent the next hour fixing what it filled in wrong.
Talking removed that trade-off. When it costs nothing to add "keep in mind the tests for this don't exist yet", I add it.
What I found when I counted
I have tics. I knew that, but not the numbers.
- "As of now": 1,575 times.
- "And so on": 1,669 times. 191 dictations end with it.
- "Let me know": 627 times.
- "Please": about once every 20,000 words.
The last one is not a manners thing, I think. It's that most of what I'm saying is an instruction to software, and you don't soften a spec.
The speech recognition has its own habits too. It hears "Claude" as "cloud" and "harness" as "hardness", which, given what I work on, comes up a lot.
What Wispr strips, and what it keeps
This is the part that changed how I write.
Wispr formats every dictation before it lands. I compared the raw speech against the formatted text for the 5,420 dictations of 20 words or more. Only 5.5% came out untouched, but the median formatted version is 97% of the original length. So it trims, it doesn't rewrite.
What it takes out is the filler. "You know" drops by 92%. "Uh" and "um" are almost all gone. "So" at the start of a sentence drops by about 39%. False starts disappear completely.
What it adds is structure. Paragraph breaks go from 10 raw dictations to 1,720 formatted ones. Colons go from 98 to 1,285. When I list things out loud, it turns them into an actual list.
And what it keeps, almost exactly, is the rest of me. "As of now" is 2.57 per thousand words when I speak and 2.63 after formatting. "And so on" barely moves. Same with "I think", "of course", "as well".
Writing is not a skill I've been blessed with, and for a long time I tried to fix that by writing more formally than I talk. Seeing this side by side, I think that was the wrong direction. My written voice is my spoken voice with the filler taken out, broken into short paragraphs. So that's how I write now, including this.
Meetings are the next part
Dictation covers what I say to my computer. It doesn't cover what gets said to me, and that's where a lot of my action items actually come from.
The setup I'm moving to has Wispr running in the background during calls, recording two tracks: my mic and the system audio. Those get merged into one transcript I can share, and an agent goes through it for action items and files them into my tracker.
The reason for two tracks is fairly simple. A single mixed recording doesn't separate what I said from what everyone else said, and for action items that's the whole question. Who said "I'll do it"? With my mic on one track and everyone else on the other, the split is already there.
It doesn't always work. I've had calls where only my mic got recorded and there was no system audio at all, so half the conversation just wasn't there. As of now I'd call this half built.
The number I don't have
Here's the admission. The question I keep asking myself is how much time all of this actually saves me. In meetings, in writing briefs, in rework.
I don't know. I've started that exact sentence in dictation and trailed off before getting to a number. 613,000 words tells you I use it. It doesn't tell you it's worth it.
That's the next thing to measure. I'll write it up when I have it.