AI-based real-time piano performance scoring on embedded systems using MIDI alignment

Music education subjects are highly computationally expensive and often lack the ability to provide real-time feedback because of subjective evaluation and high computational costs. This paper will postulate an integrated real-time pipeline that will be used to evaluate student piano play based on MIDI-based correspondence between reference scores and played sequences. The approach splits the training data into pieces using the ASAP dataset by strictly splitting the train/ test (composer, title) and samples the grid-space of possible configurations to identify the best. These are then only evaluated on the held-out test set. The streaming simulator has hard time limits and calculates lightweight online scores (unaligned note score, rhythm score, and fluency), with Direct Time Warping (DTW) alignment being performed once at the end to save spending on runtime costs, periodic DTW alignment can be optionally used to support periodic online tracking. The results of the experiment are a best mean aligned score of 21.892 and a 95% confidence interval of 19.917, 23.925 , mean step latency of 31.39 ms ( p95 40.03 ms ), 0.0000% deadline miss-rate and 90.43% alignment success on the test split. Compared with the non-aligned greedy onset–pitch matching baseline, the proposed pitch-penalized DTW alignment increased the mean note score from 5.468 to 21.892, corresponding to an absolute gain of 16.424 points, while still satisfying the real-time latency constraint. These results show that a protocol that can be real-time-assessed with accuracy and low latency can be used to evaluate the performance of the piano, it is scaleable and formative and opens up possibilities in adaptive tutoring and realizing hardware-constrained optimizations.