Polyphonic music note onset detection using semi-supervised learning

Abstract

Automatic note onset detection is particularly difficult in orchestral music (and polyphonic music in general). Machine learning offers one promising approach, but it is limited by the availability of labeled training data. Score-toaudio alignment, however, offers an economical way to locate onsets in recorded audio, and score data is freely available for many orchestral works in the form of standard MIDI files. Thus, large amounts of training data can be generated quickly, but it is limited by the accuracy of the alignment, which in turn is ultimately related to the problem of onset detection. Semi-supervised or bootstrapping techniques can be used to iteratively refine both onset detection functions and the data used to train the functions. We show that this approach can be used to improve and adapt a general purpose onset detection algorithm for use with orchestral music. ©2007 Austrian Computer Society (OCG).