Skip to content
Scott McQueen

TruSpeak · My own tool · Voice dictation

Private voice dictation that learns the words I use.

My own tool, used all day

1. Outcome

Voice dictation that runs entirely on my Mac and learns the names and terms I use every day.

2. The problem

Cloud dictation sends every word off the machine, and it still mishears client names, product names and jargon.

3. My responsibility

Product owner and daily user. I directed the build with Claude Code.

4. What was built

  • Hold a key or say a wake phrase and talk. The text appears in the window I'm working in.
  • Whisper running on the Mac's GPU, followed by my own vocabulary corrections.
  • A nightly learner that finds names it keeps mishearing in my own dictations and fixes them.
  • A morning check that the learner is learning, not just running.
  • Background jobs that step aside the moment I start talking.

5. AI and controls

  • Whisper large-v3-turbo on Apple Silicon (MLX). Nothing leaves the Mac.
  • A small Whisper model listens for the wake phrase.
  • Local text-to-speech (Kokoro) reads text back.

6. Evidence

3,987

dictations in the last 30 days

Source: dictation log, 1 Oct 2026

272,661

words dictated since 27 June 2026

Source: dictation log, 1 Oct 2026

0.55%

of dictations undone in the last 30 days (22 of 3,987)

Source: dictation log, 1 Oct 2026

2,167

automated tests, 462 of them on real recorded audio

Source: test suite, 1 Oct 2026

7. Technical implementation

Python, MLX Whisper, Kokoro ONNX, launchd

NextLuminScriptAI-assisted book writing from conversation, typing, documents and handwriting.