For language teachers
A phoneme score is not a grade. It is a location.
The engine does not say “good accent.” It scores each sound. Phone PCC 0.682 is an agreement figure at the phone level — the same unit you are reading. A word at 50 or below almost always has one or more phonemes in 0–39 or 40–59. Open that word. The red cell is the lesson.
The four bands we actually use
This is the scoring guide the engine ships with — not a rubric invented for a landing page. Native-like is 80–100. Acceptable, minor deviation is 60–79. Noticeable accent or error is 40–59. Significant mispronunciation is 0–39.
A teacher who circles a whole word is working one layer too high. The word score is a summary. The phoneme is the place the mouth went wrong. IPA is what you write in the notebook; ARPAbet is what the JSON field `arpabet` carries next to `phoneme`.
Why a word at 50 is a pointer, not a verdict
The scoring guide is blunt: words below 50 typically have one or more mispronounced phonemes. That is the whole point of going below the word. If weather is weak, you do not assign “weather” as a word. You assign DH /ð/ — or whatever cell actually dropped.
Phone PCC 0.682 versus human consensus 0.555 is computed at this grain. Two human raters, scoring phones, agree at 0.555. The engine’s phone scores sit closer to that consensus than the two humans sit to each other. That is why you can trust the red cell more than a colleague’s “I think it was the vowel.”
How to read the tape the free checker shows
The free checker paints the word tape (overallScore plus a colour per word: 80–100 clear, 50–79 close, below 50 needs work). The same response carries `words[].phonemes[]` with IPA and ARPAbet. When a word is red, that array is where you look — not at a second listen of the whole file.