Pitch Estimation of Solo and Duet Songs using Double Comb Filters

Abstract

This paper proposes a new method for the pitch estimation of sung songs for transcription. The previous pitch estimation methods that are considered most for musical instrument sounds but few for sung songs are based on the extraction of the pitch frequencies. But the extraction of the pitches is rather difficult and sophisticated signal processing is necessary since the musical sound has many harmonic components. The principle of our method is the elimination of the pitch and its harmonic frequencies. This can be performed simply by a comb filter. To adapt the fluctuation of song pitches, we use double comb filters. Using two cascade connected twelve double comb filters that are connected in parllel and are corresponding to each tone in one octave, we can estimate the pitches of solo and duet songs.

Estimating the pitch of polyphonic vocal music using comb filters connected in parallel and singular value decomposition

“…3 whereby the comb filters are arranged in a tree formation that is able to correspond to all possible harmonic combinations (the tree method [14]). We note that H 4,p d (z) is the double comb filter proposed by the authors and exhibits superior performance in eliminating fluctuating frequency components as compared to H 4,p (z).…”

“…In research into the automatic scoring of music in which musical scores are generated automatically using computers given a musical signal, estimating the pitch of the musical signal is the most important procedure and in recent years has received much attention among researchers [1][2][3][4][5][6][7][8][9][10][11][12][13][14]. The main object of analysis in this area has been the sound of musical instruments, with much active research being conducted into the estimation of the pitch of mixed signals (instruments playing at the same time) and the identification of sound sources.…”

“…On the other hand, there has been little work reporting on analysis of singing (below we assume multipart vocal singing) and so far no effective method for the estimation of pitch for polyphonic vocal music has been proposed. Against this background the authors have been conducting investigations into methods to enable the estimation of pitch for polyphonic vocal music [13,14].…”

“…. , B [11][12][13][14]. Since the periodic zero points of these H q,p (z p ) filters are able to eliminate all the frequency components of the sounds to which the filter corresponds, it is possible to estimate the pitch by detecting the zeros in the output of these.…”

“…The source of these problems lies in the increase in the number of vertically connected filters H q,p (z p ) and the fact that due to the nonstationary characteristics of the acoustics of singing, it is difficult to detect the zero points in the output of the latter components of these vertically connected bank of filters. To deal with this problem we have proposed a formation that attempts to reduce the number of filters vertically by arranging them in a tree structure so as to adapt to all harmonic combinations (below the tree method [14]); however, this approach has the disadvantage that as the number of input sounds increases, the size of the system grows dramatically. The underlying source of the problems with both methods is the attempt to estimate pitch by detecting zero points of the H q,p (z p ).…”