The proof
Phone PCC 0.682 versus human consensus 0.555.
PCC — Pronunciation Correlation Coefficient — measures how closely automatic phone scores match the consensus of trained human raters. On the public test set we use, two trained humans agree with each other at 0.555 (phone level). The production engine that scores the free checker is at 0.682 against that same consensus.
| Rater | Phone PCC |
|---|---|
| Production engine (free checker) | 0.682 |
| Two trained humans with each other | 0.555 |
What the number is, and what it is not
It is not “the AI is better than a teacher.” A teacher still decides the lesson. It is: at the phoneme grain, the engine’s scores sit closer to a trained panel than two members of that panel sit to each other. That is the only reason a red cell on DH /ð/ can replace a second listen as the default.
Phone-level is the point. Word scores and sentenceScore are different constructs (a word is a bundle of phones; a sentence also measures fluency). Do not paste 0.682 next to a fluency claim. Paste it next to a phoneme claim.
How a class should use it — and when to put it down
Use PCC as permission to sort a class by phonemes instead of by ear, on matched takes: same phrase, SNR at least 15 dB, confidence at least 0.5. Put it down when the engine itself warns — `low_snr`, `high_wer`, `low_confidence_reject`. A proof number that you apply to kitchen noise is no longer the proof number.
The bands remain the scoring guide, not PCC: 80–100 native-like, 60–79 acceptable, 40–59 noticeable, 0–39 significant. PCC tells you the tape is worth reading. The tape tells you which phone to teach.
Why this page exists
Every other page in this series that prints 0.682 or 0.555 links here. There is no third figure waiting offstage. If a paragraph needs a different number, it does not belong on these pages until that number has a source of its own.