Glossary

Mispronunciation detection

Also known as: MDD — mispronunciation detection and diagnosis

Mispronunciation detection is the task of automatically identifying which specific sounds in a spoken utterance were produced incorrectly. Its harder sibling, usually written mispronunciation detection and diagnosis, adds a second requirement: not just flagging that a sound was wrong, but saying what was produced instead — whether the sound was substituted with a different one, omitted entirely, or distorted.

Detection versus diagnosis

Detection is a binary question about each sound: right or wrong. Diagnosis asks what kind of error it was and what the speaker actually produced. Almost all commercial pronunciation APIs do detection; very few do diagnosis, and the ones that attempt it generally return a single "closest sound" field rather than a clinical error classification.

The distinction matters enormously in a therapy context. "The R was wrong" is a scoring result. "The R came out as a W" is clinically actionable information — it identifies gliding, which points at a specific treatment approach. A 2026 scoping review of thirty papers in this area found that no comprehensive system exists and that almost none classify clinical error types, with most work stopping at binary disorder-present or disorder-absent.

How mispronunciation detection is measured — and why accuracy is the wrong metric

Correctly produced sounds massively outnumber incorrect ones in most datasets. In our own benchmark, only 1.45% of phones were labelled incorrect by human judges — which means a system that flags absolutely nothing scores 98.6% "accurate".

For that reason serious evaluations report precision, recall, and F1 on the error class, plus threshold-free measures like ROC-AUC and average precision. When a vendor publishes a single accuracy percentage without stating the base rate of errors in their test set, that number is close to uninterpretable.

The other number that deserves as much attention as recall, and rarely gets it, is the false-flag rate on correct productions — how often a child who said the sound right is told they said it wrong. In a children's product that is the number that determines whether the experience is encouraging or demoralising.

What current systems can and cannot do

Omissions are the most tractable error type: a deleted sound leaves no acoustic evidence over the span where it was expected, which is a strong and unambiguous signal.

Substitutions are harder than they look. On our own benchmark the engine detected 3 of the 19 labelled substitutions in the set — a result nobody should be pleased with, on a base of 19.

Distortions are effectively unsolved. Any scorer that works by comparing phonemes against other phonemes has nothing to say about a production that is a badly formed version of the right phoneme rather than a clean instance of a different one.

The summary of where the field is: automatic mispronunciation detection is good enough to rank and triage — to tell a clinician which recordings to listen to first — and is not good enough to run unsupervised. Our own evaluation concluded that no system we measured, ours included, should be run without a human in the loop at the operating points we measured.

Mispronunciation detection — questions

What is mispronunciation detection?
Mispronunciation detection is the automatic identification of which sounds in a spoken utterance were produced incorrectly. When a system also reports what was produced instead — a substitution, an omission, or a distortion — the task is usually called mispronunciation detection and diagnosis.
How accurate is automatic mispronunciation detection?
It depends entirely on the error rate in the test set, which is why a single accuracy figure is not meaningful. On our own benchmark, where 1.45% of phones were labelled wrong by human judges, our fine-tuned engine reached a ROC-AUC of 0.901 and a best F1 of 0.325 on the error class, against 0.894 and 0.283 for the commercial API we compared with. At those precision levels most flags are false positives, which is why we describe the output as a triage signal rather than a verdict.
Why is detecting a substitution harder than detecting an omission?
An omission leaves a clear acoustic gap where a sound was expected. A substitution replaces one plausible speech sound with another plausible speech sound, so the audio still looks like ordinary speech to the model — it is only wrong relative to the target. In our benchmark both our engine and the commercial API detected the same 3 of 19 labelled substitutions.
Can mispronunciation detection replace a speech-language pathologist?
No, and no responsible vendor in this field claims otherwise. In our evaluation on real disordered child speech, the productions our engine scored as clean still included 112 that a clinician judged wrong — roughly a third of all the errors. A clinician who treated the clean bucket as cleared would miss those. The defensible use is ranking and triage inside a clinician's workflow, never a screening pass that declares anything clear.

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.

Phoneme scoring

Phoneme scoring returns a separate pronunciation score for each individual sound in a word, rather than one score for the whole word. Definition, why it matters clinically, and what to look for in an API.

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.