← Back to work

Case file 02 · Tooling · macOS

LocalFlow

Push-to-talk dictation for macOS that never leaves the laptop. Hold a key, speak, let go, and clean text appears at the cursor in whatever app you're in.

Year
2026
Role
Solo: design and build
Read
4 min
Live
Private repository · shipped
800Tests
0Words sent off the Mac
$15/moSubscription replaced
3.4sWait on a long dictation, from 11 to 21s

Context

Push-to-talk dictation for macOS that never leaves the laptop. Hold a key, speak, let go, and clean text appears at the cursor in whatever app you're in.

I built it because the tool I was paying for was cloud-only. Every word I dictated went to someone else's servers and, more to the point, it cost fifteen dollars a month. When I started looking into it, I found out very quickly that the same job could be done locally, on my own laptop.

Insights, worked out on the machine

The problem

Running everything locally means the model has to be small enough to be fast on a laptop and good enough to be worth using. Dictation is unforgiving: a wrong word takes longer to fix than it saved.

It also has to disappear. A dictation tool that steals focus, or needs its own window open, has already failed. You're dictating into something else, and the app's own UI is the one thing that must never get in the way.

The other problem only showed up once it was built: it was slow. After a long dictation, 40 seconds or more of speech, the text took another 11 to 21 seconds to appear once I let go of the key. That's too long to be useful.

What I built

A menu-bar app in Python. Speech recognition runs on the laptop through Parakeet on Apple Silicon. The text is then cleaned by fixed rules written in code, not by AI. They strip out the ums and the repeated words, write times and money as figures, and build a numbered list when you dictate one. A personal dictionary corrects the words it keeps getting wrong. After that comes an optional tidy-up by a small language model, which also runs on the laptop. It's off by default, and I run mine with S1-mini by Superwhisper, a model trained only to clean up dictation. Text goes in through the clipboard and a simulated paste, which works in every app, not just the ones that cooperate.

Around that: a transcript history in a local database, snippets that expand a short phrase into something longer (say “my LinkedIn” and it types my LinkedIn URL), per-app settings, so the tidy-up can be switched off in a terminal, and an insights pane that works everything out from your own history on your own machine. The on-screen indicator is a panel that never takes focus, so it can show what's happening without pulling you out of what you're typing into.

Transcript history, local
Snippets

Decisions & iterations

The language-model tidy-up is off by default and skipped completely for anything under five words, because most of what I dictate is a word or two and there's nothing in “Yeah.” to clean. Its replies are checked, not trusted. If a reply contains words I never said, it's thrown away and the rule-cleaned text is used. A cleanup step that can make things up is worse than no cleanup at all.

Auto-learn, which picks up my corrections for the dictionary, had to be redesigned around things people actually do. The first idea was to read the text box and spot my edits, but macOS gives no reliable way to read another app's text field, and it simply didn't work. The next version learned when I copied the whole corrected sentence, which worked but is something nobody does. Now it learns when I copy just the fixed word, or say the correction out loud, and it confirms each one on screen. Hands-free became a lock on push-to-talk instead of a separate mode, which got rid of a whole class of bugs. Auto-stop on silence was considered and rejected, because it cuts people off mid-thought.

The most useful fix was the most recent one: the wait. I measured where the time was going, made the tidy-up keep my wording and only delete and punctuate, and had both models loaded and warm before I finish speaking. That took a 50-second dictation from 16 to 21 seconds down to about 6. Then I swapped the tidy-up model for S1-mini, which is a fifth of the size and built only for this job. Long dictations now appear 2.6 to 4.3 seconds after I let go, and a typical one takes about 2.5. The on-screen pill changed too. It now says what it's doing (transcribing, removing fillers, adding punctuation, tidying up) and shows a tick once the text has gone in.

Per-app settings: tidy-up off in Terminal
Personal dictionary

Outcome

It's what I dictate with. 800 tests cover it, including the insights chart, which is tested by running the real UI file and checking the bars it actually drew. It replaces a fifteen-dollar-a-month subscription, and the audio has never left the machine.

Credits & tools
Solo: design and build — Alfie ReesPython · Parakeet MLX · S1-mini · rumpsPrivate repository · shipped
Case file 02 of 07