Patrick Donnelly, Assistant Professor
Computer Science
California State University, Chico

Abstract
Musical instrument classification is an important task in the area of Music Information Retrieval. While there have been many approaches to recognize individual instruments, the majority of these are not extensible to the more complex case of identifying the musical instruments present in polyphonic mixtures. This talk presents a data-driven clustering technique for learning regions of spectral prominence in an instrument's timbre, exploiting these regions as spectral filters in the feature extraction stage of a binary-relevance classification task. These spectral filters are used for source separation estimation in the task of automatic identification of musical instruments present in polyphonic mixtures. Machine learning experiments demonstrate this approach over several large datasets consisting of multiple articulations, dynamics, and performers, and cross-validated across datasets.

Biography
Patrick J. Donnelly is an Assistant Professor in Computer Science at California State University, Chico. Patrick received a BS in Computer Science and an AB in Music History and Italian from Washington University in St. Louis. He holds an MSE in Computer Science from the Johns Hopkins University and an MM in Musicology and an MM in Computer Music from the Peabody Conservatory at the Johns Hopkins University. Donnelly received his Ph.D. in Computer Science from Montana State University where he focused on supervised machine. His dissertation focused on multi-label classification of musical instrument timbre from polyphonic audio. As a Postdoctoral Researcher in the Emotive Computing Laboratory at the University of Notre Dame he automatically analyzed teachers' instructional practices using acoustic signal processing and machine learning. Patrick continues this research at Chico State with funding from the National Science Foundation. His primary teaching interests are in Artificial Intelligence, Machine Learning, and Programming Languages. His primary research interests are deep learning from audio, educational data mining, large imbalanced datasets, multi-label classification, and machine learning in the musical domain.