A Classification Approach to Multipitch Analysis

This paper proposes a pattern classification approach to detecting the pitches of multiple simultaneous sounds. In order to deal with the octave ambiguity in pitch estimation, a statistical classifier is trained which observes the value of a detection function both at the position of a candidate pitch period and at its integer multiples and submultiples, in order to decide whether the candidate period should be accepted or rejected. The method improved significantly over a reference method in simulations.