How we know
it works.

Where the hearing gets hard

The average share of words the microphone gets wrong, by recording condition. Clean human speech is the number to trust. The noisy room and the Bluetooth headset are where we currently do worst, and they are on this page for exactly that reason.

How each version scored

Every stable and beta build grades itself as it is published, against this same suite. So the question is not how some commit did, it is how the copy you are running did.

No released build has been graded yet. From the next release on, every stable and beta build runs this suite as it is published and its score appears here.

What happens to one clip

Every case rides the same chain the app uses, then is graded at the end.

1. 01

   Speak

   Real recorded human speech, plus a synthetic voice, a noisy-room mix and a Bluetooth-headset mix of every written case.

2. 02

   Transcribe

   The exact speech model the app ships, running on a Mac. No stand-in, no cloud, no second implementation.

3. 03

   Clean up

   The shipped cleanup pass runs inside the real app, behind the same faithfulness guard your copy has.

4. 04

   Grade

   Keyword rules, word error against the reference, and the guard's verdict. Every failure is attributed to hearing or to cleanup.

Is it getting better?

Share of cases that pass, run by run, oldest first. Typed and spoken cases are drawn apart on purpose: a spoken case can fail because the microphone misheard it, so one line across both would hide which half moved.

Every run

Each row is one full pass of the suite against one build. Open a row to read every case in it, marked up.