Cognitive and skill acquisition trajectories in school-based music education: evidence from Chinese classrooms

School-based music education in China constitutes one of the largest deployments of vocal training in the world, yet the skill-acquisition trajectories of learners in such settings, as reflected in note-level singing accuracy, have rarely been quantified at scale. Drawing on a corpus of 2,432 student-song practice sequences contributed by 1,074 distinct Chinese learners and 2,458,825 pitched note attempts derived from the publicly released singKT dataset, this study investigates how note-level correctness evolves with practice repetition, varies across vocal registers, and reflects latent learner ability. Three modeling families were compared on an 80/20 within-sequence holdout: a global mean baseline, a per-pitch baseline, a one-parameter logistic Item Response Theory (IRT) model, and a per-pitch Bayesian Knowledge Tracing (BKT) model. The analysis revealed a clear skill-acquisition curve in which mean correctness rose from 0.656 on first practice to 0.727 by the thirtieth repetition, a register effect by which low pitches were sung more accurately than high pitches (0.749 versus 0.703), and a wide spread of latent ability (theta from -3.79 to 3.36 logits). The IRT model attained an AUC of 0.673 and the BKT model an accuracy of 0.752 on held-out attempts, both exceeding the baselines by margins whose bootstrap 95% confidence intervals excluded zero, although the absolute gains were modest. Because the analysis is observational, the findings are advanced as evidence-informed hypotheses for register-graded curriculum design, repetition-rich practice scheduling and per-learner adaptive feedback that require confirmation in controlled classroom studies, and they offer a benchmark against which future deep knowledge tracing models for vocal skill may be evaluated.