For language teachers
Overall 100 and sentence 87 can both be true.
The scoring guide’s own native-speaker example of “The quick brown fox jumps over the lazy dog” comes back overallScore 100, sentenceScore 87, confidence 0.95. That is not a bug. Overall is a confidence-weighted composite of pronunciation accuracy. Sentence also scores fluency, pacing, and connected speech. Phone PCC 0.682 versus 0.555 lives at the phone layer, not at the sentence layer.
| Layer | Job |
|---|---|
| overallScore | Confidence-weighted composite of pronunciation accuracy. |
| sentenceScore | Fluency, pacing, connected speech. Good fluency above 70. |
| word | Independent per word. Below 50: open the phonemes. |
| phoneme | IPA + ARPAbet. The unit Phone PCC is computed on. |
The native take that is not “a hundred everywhere”
If you only show a student overallScore, you hide the layer that caught the pacing. If you only show sentenceScore, you hide the phone that actually failed. The four numbers exist because they disagree in useful ways. The fox sentence is the documented case: overall 100, sentence 87, confidence 0.95, no warnings.
Sentence is fluency. Word is the finger. Phoneme is the joint.
Sentence score evaluates pacing, rhythm, and connected speech. The guide’s line: above 70 indicates good fluency. A student can land every phone and still drag sentenceScore by rushing or by pausing like a list. That is a different homework from DH /ð/.
Word scores are independent. Below 50 typically means one or more phonemes in 0–39 or 40–59. Phoneme scores use the bands 80–100 / 60–79 / 40–59 / 0–39, in IPA and ARPAbet. Overall is the composite, and it is penalised when confidence is low — which is why a dirty take should never be averaged into a term mark.
PCC is a phone number, not a sentence number
When we say Phone PCC 0.682 against human consensus 0.555, we are talking about the phoneme layer. Do not quote that figure as proof that sentenceScore is “more consistent than two teachers.” SentenceScore is a different construct (fluency). Use PCC to trust the red cell. Use sentenceScore to decide whether the problem is a sound or a stretch of speech.