Automatic Musical Instrument Identification

Joseph McKay (Thesis), Dublin Institute of Technology

Document Type Theses, Masters

Successfully submitted for the award of Master of Philosophy (M.Phil) to the Dublin Institute of Technology, October, 2011.

Abstract

The increased availability of digital repositories of music coupled with the increase in demands of Music Information Retrieval (MIR) has spawned a growth in content-based multimedia retrieval systems. On a commercial level consumers demand more intelligent systems to satisfy their digital music requirements. Areas of MIR, which includes musical transcription, similarity, segmentation and classification would benefit significantly from efficient musical instrument classification technology. Many instrument classification systems have attempted to meet this challenge, yet most state-of-the-art systems are restricted to the monotimbral classification task. When faced with the real-world scenario of classifying polytimbral stereo musical mixtures, these systems lack the capability of performing the sophisticated processing required. Research has identified as desirable the reduction of the polytimbral classification task to that of classification of individual instrument sources which have been separated prior to the instrument classification process by way of sound source separation (SSS) algorithms. Such a system would prove powerful in MIR. The principal aim of this research is to establish a ground truth of the capabilities of the SSS algorithm ADRess as a preprocess to instrument identification in polytimbral musical signals. A thorough review of the instrument identification literature is provided which identifies the need for the capabilities of ADRess. Most popular digital music is in stereo format. Ground truth experiments using basic acoustic features with two common classifiers, k-Nearest Neighbour and Gaussian Mixture Model, succeed in identifying the ADRess separated sources from a synthetic polytimbral mixture of five non-percussive instruments. This novel contribution to the field of musical instrument identification opens the door to new possibilities/alternatives in the classification of stereo polytimbral signals. Polytimbral classification has been reduced to a simpler task. Current state-of-the-art systems can use ADRess as a preprocessing step and continue with their current timbre modelling approaches. However, ADRess does have its limitations. It is subject to artifacts distorting classification findings as a result of resynthesis of sources with overlapping harmonic partials. It is also at the mercy of the sound engineers who choose to position instruments across the stereo spectrum. However, instrument identification is no longer subject to the restrictions of current polytimbral approaches and ADRess o ers a better alternative to SSS as a preprocessing step than current algorithms such as Independent Component Analysis and DUET.