Yuriy Balaka

Say in Place —
voice typing for Mac

What was wrong

I speak Russian with English terms in the same sentence: “send me the updated flow before the demo”. I wanted to type by voice and get exactly what I said. Not a translation. Not a version polished by an AI.

What I decided

  • The app types what you said. It does not rewrite it. It removes only what the recogniser invents or repeats, and the replacements you set in your own dictionary.
  • Speech is recognised on the Mac. The audio is deleted after recognition.
  • “Nothing leaves your Mac” has to be true in the code, not only in the privacy policy. Before release I had the code checked line by line. Three things made the promise false. The worst: with one setting on, the full audio went to an outside server. I removed all three before the name Say in Place existed.
  • The app uses the network in exactly three places: downloading the speech model, checking for updates, and the purchase page, which opens in your browser.
  • I tested 12 speech models on 174 of my own dictations. The fast ones made several times more mistakes on mixed Russian and English. I kept the model the app already used and made it faster instead.

What I built

A Mac app. Hold Fn, speak, release: your words appear where the cursor is, in a messenger, in email, in a chat with an AI, in a code editor. While you hold the key, the app cuts your speech at the pauses and recognises the finished pieces. When you let go, only the last piece is left. So the wait after release is about the same for a short note and a long one.

I built it alone, with AI coding agents (Claude). I set the tasks, made the decisions, and checked the result by eye and by ear. The agents wrote the code, ran the measurements and wrote the reports.

What changed

  • The first working prototype ran on 11 April 2026.
  • In the first 15 days I dictated 4,960 times. That is about 36 hours of speech.
  • Version 1.0 came out on 26 September 2026. The site sayinplace.com opened, and sales opened the same day. You pay once.

What is left

  • Long dictations still have punctuation problems where the pieces join. The site does not promise clean long text until this is fixed.
  • Mixed languages are measured mainly on Russian with English.