Note-Based Segmentation and Hierarchy in the Classification of Digital Musical Instruments

The ability to automatically identify the musical instruments occurring in a recorded piece of music has important uses for various music-related applications. This paper examines the case of instrument classification where the raw data consists of musical phrases performed on digital instruments from eight instrument families. We compare the use of extracted features from a continuous sample of approximately one second, to the use of a systematic segmentation of the audio on note boundaries and using multiple, aligned note samples as input to classifiers. The accuracy of the segmented approach was greater than the one of the unsegmented approach. The best method was using a two-tiered hierarchical method which performed slightly better than the single-tiered flat approach. The best performing instrument category was woodwind, with an accuracy of 94% for the segmented approach, using the Bayesian network classifier. Distinguishing different types of pianos was difficult for all classifiers, with the segmented approach yielding an accuracy of 56%. For humans, broadly similar results were found, in that pianos were difficult to distinguish, along with woodwind and solo string instruments. However there was no symmetry between human comparisons of identical instruments and different instruments, with half of the broad instrument categories having widely different accuracies for the two cases.