Dictation that keeps up with speech.

I was paying $15 a month for a cloud dictation app with a one-to-two-second delay between pressing the key and the app actually listening. That delay broke my flow. Investigating it showed why: the microphone opened cold on every keypress, Bluetooth headsets added a profile switch on top of that, and audio passed through several processes before transcription.

I built it solo with Claude Code. I benchmarked speech-recognition and cleanup models on my own voice before building, then built the native Swift app, the audio pipeline, the fine-tuning data and run, and the evaluation. Along the way I diagnosed and fixed the recognizer outputting “Thank you.” on trailing silence, fixed an MLX-to-GGUF convolution-weight layout incompatibility for LFM2, and set up a self-signed signing identity so macOS privacy permissions survive rebuilds.

1. 01Run as one native Swift menu-bar process. The microphone stays open in a ring buffer with 0.3 s of pre-roll, and a listen-only event tap watches the fn key, so recording starts about 50 ms after key-down.
2. 02Segment audio at pauses and transcribe while I am still talking, using Parakeet TDT 0.6B v3 on the Apple Neural Engine through FluidAudio and CoreML, so little work is left at release.
3. 03Fine-tune a small cleanup model instead of prompting a larger one: LFM2.5 1.2B Instruct, LoRA-trained with MLX on about 3,700 of my own raw-to-cleaned dictation pairs and run through llama.cpp on Metal.
4. 04Read the surrounding text through Accessibility APIs so dictation can continue mid-sentence in lowercase, fix spacing, and spell names that are visible on screen correctly. Paste, then restore the clipboard.
5. 05Use llama.cpp’s prebuilt framework rather than MLX-Swift so the app builds and ships with Command Line Tools only, without Xcode.

On an M5 MacBook Air, across 130 held-out real recordings (44 minutes): recording starts about 50 ms after key-down, speech recognition takes 0.09 s median for clips up to 20 s, and text is pasted about 250 ms after release. Word error rate was 3.2% against the cloud service’s own transcripts; Whisper large-v3-turbo scored 8.0% and was three times slower. The fine-tuned 1.2B cleanup model scored 10.5% content WER against 15.0% for a prompted Qwen3 4B, at under half the latency, with no dropped or invented content in testing. The app also includes searchable history, a personal dictionary, settings, and one-click history import from the cloud app.

These results come from one speaker on one Mac. Error rates are measured against the cloud service’s transcripts rather than hand-checked references, and the cleanup result reflects the held-out test set, not a guarantee for every dictation.

Evaluate across more speakers, microphones, and Macs, and add a hand-checked reference set so error rates do not depend on another service’s transcripts.