How speech clarity is measured here.

What the number actually comes from

Nothing is scored until the recording is checked. Is there any signal at all, is it long enough, is the room louder than the speaker, did the microphone clip. A recording that fails those checks is handed back for another attempt and costs you nothing, because scoring noise would produce a number that means nothing.

What survives that check goes to a phoneme model, which returns a confidence from 0 to 100 that the target sound was produced — per sound, not per attempt. A single number for a whole word tells you something went wrong without telling you where; a score attached to a specific sound in a specific position is something you can act on. Alongside it sit the acoustic measurements the judgement rests on: voice onset time, pitch, aspiration, vowel shape.

Every score is paired with a written cue in plain language. That is deliberate: a learner who cannot hear a played-back recording must still be able to complete the feedback loop, so if something matters it is on the screen as text, never as audio you are expected to judge by ear.

And the honest limits. The score is a practice signal, not a clinical measurement. It moves with microphone quality, background noise and accent. It does not diagnose anything, it does not measure hearing, and it is not an assessment — a therapist’s judgement is a different thing entirely, and the app is built to sit alongside one rather than replace it.