Pitch estimation of polyphonic songs using comb filters and singular value decomposition

Abstract

A new pitch estimation method for song sounds using twelve comb filters and singular value decomposition (SVD) processing is presented. This method is based on the ability of a comb filter, corresponding to one tone, to eliminate the fundamental frequency and other harmonic frequencies, and also on the relationship between the number of singular values obtained from the SVD and the number of the signal harmonic components. The pitches can be estimated by using a set of twelve comb filters connected in parallel to eliminate one tone in the input, and by identifying outputs with a smaller number of singular values. Pitch estimation results for a duet and a trio with this method are presented.