You had six things in your head at the top of the stairs and four by the time the app opened.
The other two are not gone. They will come back at 11pm, one at a time, while you try to sleep. Between the head and the box there is a keyboard, and the keyboard is where the list gets thin. You typed "dentist", and by the time you had spelt "appointment" the thing about the car insurance had left.
The head is faster than the hands
Nobody types as fast as they worry.
A full head does not queue. Six things arrive at once, and they all want to be first. Typing takes them one at a time, one word at a time, with spelling in between, and while the hands work on the first thing the other five have to wait. With ADHD it is hard to keep a list in your head. A 2013 meta-analysis of 38 studies of adults found this well supported (Alderson and colleagues, 2013). So the waiting room is small, and the things waiting in it leave.
This is the quiet failure of the brain dump as most people do it. The idea is right: get everything out of the head and onto a page. The tool is wrong for the speed. You are trying to empty a bucket through a straw.
Speaking is the bucket tipped over.
A voice dump is a brain dump you do not have to hold still for
Talk, and do not edit.
Said out loud, the list comes out in the order it is in your head, which is no order at all, and no order is fine. "The dentist, and the thing with the car, the insurance, it's due Friday, or Thursday, and I owe Sam a reply, and the plants." Half sentences. A pause while you look at the ceiling. A second pause while the next thing surfaces. Nothing about this would survive a keyboard, because at "insurance" you would stop to spell it and at "Friday" you would stop to check, and by then the plants would be gone.
Out loud, the plants make it. The speed of the mouth is close enough to the speed of the head, and there is no second job in the way. No spelling, no ordering, no deciding which line the dentist goes on. Those jobs still need doing. They are done after, by something with a steadier head than yours at 6pm.
The only rule is the one you already know from a written dump: everything goes in, and nothing gets sorted while it is coming out.
What we measured: 28 speech models, 39 minutes of rambling
Most speech models are built for clean speech, and a brain dump is not clean.
Before we picked the one we ship, we ran 28 speech models on two kinds of audio. Short, tidy clips, the kind most models are tested on. Then 39 minutes of long, rambling dumps with pauses up to 7 seconds, the kind a real person makes at the end of a day. The clean clips told us little. The rambles told us everything.
Chat-style audio models, the ones built to talk back, made things up over silence. Given seven seconds of nothing, they filled it. One answered the clip instead of writing it down. The newer version of the model we ended up with, Parakeet v3, dropped quiet speech: a quiet line vanished from the transcript with no mark left behind. We called these silent drops, and on a to-do list a silent drop is the worst error there is. A wrong word is a typo you spot. A dropped sentence is a task you never see again.
We ship Parakeet TDT 0.6B v2. On our measure it got 7.9% of words wrong across everything, 1.8% on the dumps, and zero silent drops. Not the newest model and not the most talkative. The one which wrote down what you said, including the quiet bit at the end.
How Kaapi does it
There is a microphone beside "Turn it into tasks".
Press it and talk. Listening is silent, and nothing talks back. You can talk for minutes, with pauses as long as you need, and the words land in the box as text you can read and change. A seven-second pause while you stare at the ceiling is fine; nothing fills it and nothing is dropped. Then "Turn it into tasks", and a model running on your Mac turns the ramble into small tasks with steps. A time you said, "by 4pm" or "the weekend", becomes the task's deadline.
Your voice is turned into words on the Mac, and the words are turned into tasks on the Mac. Nothing you say leaves it. Why the model runs there, and what it costs you, is here.
How to do this without Kaapi
You need something to listen and a rule against tidying.
- Open the voice memo app on your phone. Press record before you have decided what to say, and talk until the head is quiet. Pauses are allowed. Repeats are allowed.
- The same day, play it back and write one line per thing. Do not merge, do not rank. "Dentist." "Car insurance, Friday." "Reply to Sam." "Plants."
- Give every line with a day in it a time of day. "Friday" becomes "Friday 4pm". A time you can see is a deadline your brain will start on.
- If a line is a project and not a task, write its first small step under it, and use the step as the task.
Frequently asked questions
Is a voice brain dump better than a written one for ADHD?
It is faster, and speed is the point. With ADHD it is hard to hold a list in your head, and typing asks you to hold it while your hands catch up. Speaking keeps closer to the speed of the head, so fewer things fall out before they are recorded. Sort them afterwards, not while you speak, and sort them the same day, while the recording still means something to you.
What if I ramble or go quiet in the middle of a voice dump?
A brain dump sounds like a ramble with gaps. Pauses while the next thing surfaces are part of it, and the tool has to cope with them. We measured 28 speech models on rambling dumps with pauses up to 7 seconds because most models do badly there, and we picked the one with zero silent drops.
Does a voice brain dump have to be sent to the cloud?
No. Speech can be turned into words on the Mac itself, which is how we do it, so nothing you say leaves your computer. The trade is one model file fetched once, and a little time on the Mac's own chip. For the rawest writing you do, the trade is worth it.
How long should a voice brain dump be?
As long as the head is full. Most are a few minutes. Stop when the next thing does not come, not when you hit a length. A long dump is fine as long as nothing gets dropped on the way to the page, which is why we measured the speech model on 39 minutes of long rambles, not on tidy clips.
Sources
- Alderson, R. M., Kasper, L. J., Hudec, K. L., and Patros, C. H. G. (2013). Attention-deficit/hyperactivity disorder (ADHD) and working memory in adults: A meta-analytic review. Neuropsychology, 27(3), 287–302. doi.org/10.1037/a0032371
- Kaapi (2026). Speech bench: 28 speech models run on short clips and on 39 minutes of long, rambling dumps with pauses up to 7 seconds. Parakeet TDT 0.6B v2: 7.9% of words wrong overall, 1.8% on the dumps, zero silent drops. Our measurement, not a trial.