Tone Perception Trainer

Tone Perception Trainer is built to the one rule that makes tone training work: nobody grades their own ear. Perception runs a real high-variability paradigm — randomized multi-talker presentation, two-alternative forced choice, immediate feedback, per-trial reaction time — and scores it as d′ and a confusion matrix rather than raw accuracy, because raw accuracy hides that you are riding a base rate. Singles are never offered: the inventory is the 20 Mandarin tone pairs (16 ordered pairs plus neutral as a second element), and the app ships the full obligatory sandhi rules (T3+T3, half-third, 一 and 不) with Chao tone-letter values, so it can show that the card says nǐhǎo while the clip says níhǎo. Japanese accent is a separate mode with the nucleus as an integer over mora count, the shipped priors as a null model (~47% heiban, ~26% antepenultimate; ~70% of native nouns unaccented vs ~7% of loanwords), and one hard rule: odaka vs heiban is never offered on an isolated word, because that contrast is inaudible without a following particle. It works with nothing installed: the bundled bank uses the speech voices already on your device, and the app states plainly how many talkers that is and whether they can support high-variability training before you invest a session. Add your own native clips any time — each is labelled once, its pitch contour measured, and clips whose measured contour differs from the citation form are surfaced honestly rather than declared wrong, because emphasis, focus, final lowering and sandhi are all legitimate. Production closes the loop in under a second: your pitch is extracted with a YIN tracker, passed through a voicing-confidence and octave-consistency gate, DTW-aligned to the reference, and both contours are drawn overlaid so you can see the tracker's work and reject it. When the gate fails the app refuses with a reason instead of guessing; when it passes you get a specific number (‘your fall on syllable 2 is 3.1 semitones