Learning spectral filters for single- and multi-label classification of musical instruments

Abstract

Musical instrument recognition is an important research task in music information retrieval. While many studies have explored the recognition of individual instruments, the field has only recently begun to explore the more difficult multi-label classification problem of identifying the musical instruments present in mixtures. This dissertation presents a novel method for feature extraction in multi-label instrument classification and makes important contributions to the domain of instrument classification and to the research area of multi-label classification. In this work, we consider the largest collection of instrument samples in the literature. We examine 13 musical instruments common to four datasets. We consider multiple performers, multiple dynamic levels, and all possible musical pitches within the range of the instruments. To the area of multi-label classification, we introduce a binary-relevance feature extraction scheme to couple with the common binary-relevance classification paradigm, allowing selection of features unique to each class label. We present a data-driven approach to learning areas of spectral prominence for each instrument and use these locations to guide our binary-relevance feature extraction. We use this approach to estimate source separation of our polyphonic mixtures. We contribute the largest study of single- and multi-label classification in musical instrument literature and demonstrate that our results track with or improve upon the results of comparable approaches. In our solo instrument classification experiments, we provide the seminal use of Bayesian classifiers in the domain and demonstrate the utility of conditional dependencies between frequency- and time-based features for the instrument classification problem. For multi-label instrument classification, we explore the question of dataset bias in a cross-validation study controlled for dataset independence. Additionally, we present a comprehensive cross-dataset study and demonstrate the generalizability of our approach. We consider the difficulty of the multi-label problem with regards to label density and cardinality and present experiments with a reduced label set, comparable to many studies in the literature, and demonstrate the efficacy of our system on this easier problem. Furthermore, we provide a comprehensive set of multi-label evaluation measures.