Transparency
Most apps hand you a number and call it an “AI examiner”. We'd rather show you how it's made — because trusting the score is half the training. Here's the engine, the method, and, plainly, where the score stops.
The engine
The acoustic layer is Azure Pronunciation Assessment running on the Canadian-French model (fr-CA), in the canadacentral region — the same pronunciation the TCF Canada tests, not European fr-FR. On top, an LLM layer reads the result against the TCF Expression Orale criteria. Both names are here on purpose: you can check what each one does.
The method
You record a phrase. The audio goes to the fr-CA engine, which scores every word — and every sound inside it — for accuracy, fluency and completeness. We map that to the 10–11 / 20 Expression Orale band (the NCLC 7 speaking level) and pull out the weakest sounds, which become your review loop. No magic number from nowhere: every colour comes from an engine value.
What the score says
The score comes in four 0–100 axes — pronunciation, accuracy, fluency, completeness. The per-word colours (green, amber, red) are the real phonetic accuracy the engine measured, not a guess. On Test-Prep, the NCLC estimate is exactly that: an estimate, calibrated to the EO 10–11 band — a compass for training, not the official result.
The limits, plainly
We score read-aloud pronunciation — not the depth of a spontaneous conversation (that's the conversation partner, and it isn't graded). An estimate is not the official France Éducation International score. And the acoustic engine has its own error, like any model. Showing that is the point: you trust what the score says because you also know what it doesn't.
No magic, no “+X guaranteed points”. Just an honest engine and a plan to get better sound by sound.