Voices in Polyphonic Music : A Contig Mapping Approach

The algorithm presented in this paper uses basic principles of music perception to separate voices in polyphonic music using contig mapping. Three metrics are proposed to evaluate performance and accuracy of the automatically separated voices. These metrics are: the average fragment consistency (AFC), the correct fragment connection rate (CFC), and the average voice consistency (AVC). In this method synchronous notes are not allowed to be part of the same voice and contig mapping shows high fragment consistency, the grouping of notes from the same voice into the same fragments. The authors use the knowledge of the perceptual principals of auditory streaming to create an O(n 2) contig mapping algorithm for separating polyphonic pieces into their component voices. The algorithm only considers pitch height and event boundaries and ignores information on timbre and sound source. The easiest method for voice separation, adopted by most sequencer software packages, is split voices according to some set of non-overlapping pitch ranges, but it can produce highly inaccurate and unsightly results. Thus, many researchers have tried to improve it. The contig mapping approach, used in this method, is based on several underlying perceptual principles. One of the principles is the " pitch proximity principle " , which means that " the coherence of an auditory stream is maintained by close pitch proximity in successive tones within the stream ". The other perceptual principle is the " stream crossing principle " , which is related to the human's difficulty in tracking streams of sounds that cross with respect to pitch. The rules and principles are translated into some assumptions implying that at certain segments of time (called maximal voice contigs), all voices sound synchronously in a well-behaved manner and also that intervals between the successive notes in a stream or voice are minimized. Based on the assumptions, distance minimizing procedure can be used to connect voices between segments. The following terms are used in description of the algorithm: note (a musical entity with pitch & duration); fragment (a sequence of successive notes belonging to the same voice); contig (a collection of overlapping fragments that the overlap depth at any time is constant); and finally, maximal voice contig (a contig with the maximum number of voices present). The algorithm is based on a segmentation procedure, in which the piece is segmented according to voice count that remains constant within the contig, and a connection policy, which is …