Machine Learning of Expressive Microtiming in Brazilian and Reggae Drumming

We have used supervised machine learning to apply microtiming to music specified only in terms of exact, quantized, theoretical note times. Specifically, the input consists of a sort of musical score: note data specifying the quantized attack times for a variety of percussion instruments. The output of the regression schemes we tried is simply the microtiming deviation to apply to each note. In particular, we trained Locally Weighted Linear Regression / K-NearestNeighbors (LWLR/KNN), Kernel Ridge Regression (KRR), and Gaussian Process Regression (GPR) on data from skilled human performance of a variety of reggae drum parts and Brazilian rhythms. Evaluating our results with cross-validation, we found that the three methods are quite comparable, and in all cases the mean squared error is substantially less than the mean squared microtiming of the original data. Subjectively, our results are satisfactory; the applied microtiming captures some element of musical style and sounds much more expressive than the quantized input.