Unsupervised Learning of Structural Representation of Percussive Audio Using a Hierarchical Dirichlet Process Hidden Markov Model

A method is proposed that extracts a structural representation of percussive audio in an unsupervised manner. It consists of two parts: 1) The input signal is segmented into blocks of approximately even duration, aligned to a metrical grid, using onset and timbre feature extraction, agglomerative single-linkage clustering, metrical regularity calculation and beat detection. 2) The approx. equal length blocks are clustered into k clusters and the resulting cluster sequence is modelled by transition probabilities between clusters. The Hierarchical Dirichlet Process Hidden Markov Model is employed to jointly estimate the optimal number of sound clusters, to cluster the blocks, and to estimate the transition probabilities between clusters. The result is a segmentation of the input into a sequence of symbols (typically corresponding to hits of hi-hat, snare, bass, cymbal, etc.) that can be evaluated using the Adjusted Random Index (ARI). As a proof-of-concept, the system segmentation has been tested using two simple Disco-style drum loops, yielding a an ARI of 56% for the best stable HDP-HMM parameter setting.