English

Preliminary report · last updated 18 July 2026

Accuracy on real recorded speech

Real human recordings replayed through the app's exact pipeline, scored word-by-word against human-made transcripts of the same audio:

Source	Style	Word accuracy
Tech commentary (fast, casual US)	Rapid conversational	97.1%
Explainer video (natural UK)	Natural narrative	95.3%
Audiobook narration × 3 narrators	Long-form prose	99–100%
Synthetic business documents × 3 voices	Email & contract dictation	97–100%

Most "misses" against captions are actually filler words the captions dropped but the engine faithfully heard. About 30 minutes of unique audio and ~7,900 words scored to date.

Speed

Text ready about 0.5 s after releasing the key.
Median ~1.0 s from key release to text pasted at the cursor (measured across real daily use, MacBook Neo).

Formatting

The deterministic formatting rules are verified by an automated suite of 235 checks that runs before every release — phone numbers, currency, thousands separators, ordinals, clock times, percentages, fractions, decimals, IP addresses, version numbers, web and email addresses, letters, and lists. AI formatting is additionally held to a word-for-word guarantee: it may rearrange, never delete or invent.

Notes

All processing is on-device; results above are from an Apple Silicon MacBook Neo (8 GB).
Comparison replays against other dictation apps use the same audio through each app's real pipeline.
This is a preliminary report; the full published methodology (including audio sources and scoring scripts) arrives with the finished benchmark round.