Tim Shuttleworth and R.G. Wilson, Note Recognition in Polyphonic Music using Neural Networks (October 1, 1993).
This report describes the initial results of work toward the development of a system for the automatic transcription of polyphonic musical works where the input to the system is a digital sound waveform representation (such as the data stored on a Compact Disc), and the output is a symbolic, score-like representation of the work being performed. A lowlevel representation of the music signal is discussed (the Multiresolution Fourier Transform). A method of finding the beat of a piece of music is described, as is some preliminary work on pitch tracking through use of the frequency centroid. One of the major differences between this work and that previously published is that is an emphasis on the use of higher level structural properties of music. Possible methods of integrating high- and low-level musical knowledge are discussed, and in particular the Hidden Markov Model paradigm, previously found in work on speech recognition, is introduced in a music recognition context. Finally there is some discussion of future directions for this work.