Real recordings, not scripts
Our test set uses hundreds of public recordings of people talking naturally, with accents, pauses, and plenty of ums.
Lab
We test it on real voices.
Changes to how SpokenKeys hears and cleans up your words are checked against hundreds of real recordings before they ship. Here's what we've measured so far.
- for the AI to clean up what you said
- ~0.3 s
- word errors with other people talking nearby, after we taught it to follow the main speaker
- 23.5% → 7.5%
- questions it answered instead of typing them, across 45 test prompts
- 0 of 45
- real English recordings in our test set
- 300+
How we test
Real speech, blind checks, lots of reruns.
Blind side-by-side checks
When we change how the AI cleans up text, we compare old and new results without knowing which is which, and keep the winner.
Hundreds of automatic checks
Each Mac release has to pass roughly 300 automated tests, run ten times in a row, before it ships.
Follow the Lab: