Sound-by-sound notes

How ArticScore Scores Individual Speech Sounds

Automatic pronunciation scoring does not behave the same way on every sound. These pages take the sounds that matter most in speech therapy — /r/, /s/, /l/ and “th” — and set out, for each one, why it is difficult for children, which of its common errors a phoneme scorer can detect, which it structurally cannot, and what ArticScore actually returns for a word containing it.

The sounds

Error type predicts performance better than sound identity

The useful way to think about these pages is not “which sounds is the engine good at” but “which errors is it able to represent”. A substitution replaces one phoneme with another phoneme the model already knows, so the engine can score the target low and name what it heard instead. An omission leaves no acoustic evidence where a sound was expected, which is a strong signal. A distortion is neither: it is a badly formed version of the right sound, it is not any other symbol in the inventory, and there is nothing for it to lose against.

Scoring by sound — questions

Does automatic scoring work equally well on every speech sound?
No. Performance varies substantially by sound and, more importantly, by error type. Sounds whose common errors are substitutions with another phoneme — "th" replaced by /f/, /l/ replaced by /w/ — are handled comparatively well, because the substitute exists in the model's inventory. Sounds whose common errors are distortions, above all /r/ and /s/, are handled poorly, because a distortion is not any other phoneme.
Which sound is hardest for a pronunciation scoring API?
/r/. It is the latest-acquired consonant in American English, it can be produced with two entirely different tongue shapes that sound the same, and its most common clinical error is a distortion rather than a substitution — the error class that phoneme-based scoring is structurally unable to see.
Can any API detect a lateral lisp?
Not by phoneme scoring. A lateral lisp is still an attempt at /s/ and is not any other phoneme in the inventory, so a closed-phone-set scorer has no way to represent it and will typically score it as correct. If detecting lateral lisps is a core requirement, no phoneme-scoring API we are aware of — ours included — will meet it.
How is this different from the /sounds/ pages on this site?
The /sounds/ pages are written for parents and children: how to make the sound, word lists, and practice activities. These pages are written for developers and clinicians evaluating automatic scoring: how the sound behaves under a phoneme scorer, which of its errors are detectable, and what the API response looks like. Each page here links across to its practice counterpart.