Glossary

Phoneme scoring

Also known as: per-phoneme pronunciation scoring

Phoneme scoring is pronunciation assessment that returns a separate score for each individual speech sound in a word, rather than a single score for the word or the sentence as a whole. Scoring the word "rabbit" at the phoneme level produces distinct results for /ɹ/, /æ/, /b/, /ɪ/ and /t/, so an application can identify precisely which sound was the problem instead of only reporting that the word was mispronounced.

Why the granularity matters

A whole-word score answers "was that right?". A phoneme score answers "which part was wrong?" — and in speech therapy, the second question is the only one that leads anywhere. Therapy targets sounds, not words. A child works on /r/ across dozens of words, and progress is measured per sound and per position within the word.

Phoneme scoring is also what makes position-specific reporting possible. The same sound can be solid in one position and unreliable in another — a child may produce /s/ correctly at the end of "bus" but not at the start of "sun" — and that distinction determines where therapy starts. A word-level score cannot express it.

For a language-learning product the calculus is different but the conclusion is similar: learners improve fastest when feedback identifies the specific segment to work on.

What good phoneme-level output contains

A score per phoneme, on a documented scale, with the calibration behaviour described rather than left as an opaque number.

The phoneme identity in a standard notation — IPA, ARPAbet, or both — so the output can be joined against clinical materials and word lists.

Time spans, so the application can highlight or replay the exact fragment.

What was heard instead, when the score is low. This is the field that separates a scoring API from an explanation API, and it is the field that is most often missing or unpopulated exactly when it would be most useful.

An error type where one can be determined — substitution, omission, or distortion — because that vocabulary is what a clinical report is written in.

The limits of scoring at this granularity

Per-phoneme scores are noisier than word-level scores, because each one is computed over a shorter span of audio. Aggregating them back up to a word score smooths that noise, which is part of why utterance-level correlations with human judges are consistently higher than phone-level ones for every system we have measured, including ours.

Phoneme scores also saturate: a large share of correct productions land at the very top of the scale, so the difference between a good production and an excellent one carries little information. Ranking productions by score works better than thresholding them.

And no phoneme score can represent something that is not a phoneme. Distortions — the productions that are a poor version of the right sound rather than a clean version of a wrong one — remain invisible to this entire class of system.

Phoneme scoring — questions

What is phoneme scoring?
Phoneme scoring is pronunciation assessment that produces a separate score for every individual sound in a word rather than one score for the whole word. It lets an application tell a user exactly which sound was mispronounced instead of only that the word was wrong.
What is the difference between phoneme scoring and word scoring?
Word scoring returns one number for the whole word. Phoneme scoring returns one number per sound within it. Word scores are less noisy and correlate slightly better with human overall judgements, but they cannot tell you which sound to work on — which is the information a speech therapy or literacy application actually needs.
Which pronunciation APIs return phoneme-level scores?
Several vendors do, including the major cloud providers' pronunciation assessment services, a handful of specialist language-learning APIs, and ArticScore. What separates them is not usually whether they return per-phoneme scores but what accompanies each score — whether the API also names the sound it heard instead, classifies the error type, returns timing spans, and whether the underlying acoustic model has been tuned for children's voices. Our buyer's guide sets out the questions to ask.
Is phoneme scoring accurate enough to use with children?
It is accurate enough to guide practice and to triage which recordings a clinician should review, and it is not accurate enough to run without one. The figure to ask any vendor for is the false-flag rate on correctly produced sounds by child speakers — how often a child who said it right is told they said it wrong. On our benchmark, our fine-tuned engine flagged 6.5% of correctly produced phones at its shipped operating point.

Related terms

Goodness of Pronunciation (GOP)

Goodness of Pronunciation (GOP) is the standard algorithm for scoring how well a speaker produced an expected phoneme. Definition, the formula, its known weaknesses, and how ArticScore uses it.

Forced alignment

Forced alignment maps a known transcript onto the audio, marking where each word and phoneme begins and ends. Definition, how it works, why it fails on child speech, and its role in ArticScore.

Mispronunciation detection

Mispronunciation detection is the task of automatically flagging which sounds a speaker produced incorrectly. Definition, why accuracy is a misleading metric, and the performance limits of current systems.