How do you measure a student’s pronunciation without sitting through every take?

Forced alignment against a phrase, not a vibe

The engine does not grade “accent” and it does not treat free transcription as a mark. You send audio plus the exact sentence the student was asked to say. Forced alignment locks each phone to that reference. If they said a different sentence, that shows up as a mismatch — decodedTranscript versus the phrase — not as a quietly inflated overallScore.

That is the difference between a pronunciation meter and a speech-to-text box with a thumbs-up. The meter only exists because the reference text is known.

Four scores, four jobs

Every assessment returns the same four 0–100 layers: overallScore (confidence-weighted composite), sentenceScore (fluency, pacing, connected speech), per-word scores, and per-phoneme scores in IPA and ARPAbet. Words below 50 almost always hide one or more phonemes in 0–39 or 40–59.

You also get confidence (0 to 1) and, when the take is dirty, warnings. A high overallScore with confidence under 0.5 is not a high score. Throw the take out and re-record.

• Same phrase for the whole class — otherwise you are not measuring the same thing.
• Sort by words below 50, not by who “sounded better.”
• Ignore any row whose confidence is under 0.5. That is the engine telling you the audio and the phrase do not match.

What this looks like in a class of twenty

Twenty files of the same sentence. You do not play twenty files. You read twenty word tapes. The students whose DH or IH collapsed are the ones who get a note tonight; the rest do not need a teacher’s ear on that homework. Phone PCC 0.682 versus 0.555 is the reason you can do that without pretending two human listeners would have agreed anyway.