Svar

I was using Wispr Flow, but transcription was slow, and I wanted things very specific to me, like pasting my last transcript into an input box that wasn't selected while I was dictating.

So I built my own, in public, in a thread on X.

A native macOS app in SwiftUI, my first time with it.

Fully offline, with a choice of local models for transcription and cleanup. I keep trying new ones; right now it's Parakeet Ultra for transcription and S1-mini by Superwhisper for cleanup.

Uses macOS's accessibility and OCR APIs to read what's on screen, so names and terms come out right.

Keeps my recordings and transcripts for fine-tuning, and as evals to test new models when I add them.

It will most likely never launch. It's a personal tool I use every day.