Feature extraction of bass guitar using principal component analysis
Y. Abe, Yuki Murakami, Masanobu Miura · 2010
Drums and bass guitars create rhythm in popular music. A drum pattern database including thousands of music excerpts was previously developed for investigating the tendency of constitution of drum in musical score. For the investigation of bass guitar, several studies have been done, such as estimating the fundamental frequencies (F0) of the melody or bass lines using monaural audio signals containing sounds from various instruments and focusing on the bass guitar to classify genres. However, the tendency for bass guitar in popular music using many patterns has not been investigated; therefore, the development of a database for bass guitar profiles, such as the rhythm and pitch, has not been reported. We propose an identification method for the MIDI tracks of the bass guitar from MIDI excerpts comprised of several instrumental tracks. The proposed method identifies the bass guitar part using a heuristic approach from several bass guitar players, so that an accumulation of patterns of bass guitar is obtained. Then, each one-bar length is identified as either a bass guitar part, called a “bass guitar pattern” or not. The onset, interval, and dynamics profiles are extracted from the bass guitar patterns. The onset profile represents the onset time sequence, the interval profile represents the interval for each note from the root note of the excerpt, and the dynamics profile represents the rank in value of the MIDI velocity for each note. We constructed a bass guitar pattern database comprised of these three types of profiles. We introduce several parameters for automatic arrangement, such as those from the proposed database, using principal component analysis (PCA). We call these parameters the “Eigenphrase of bass guitar”. Using PCA, we extract the principal components for the “Onset profile for eigenphrase of bass guitar”, “Interval profile for eigenphrase of bass guitar”, and “Dynamics profile for eigenphrase of bass guitar” from the onset, interval, and dynamics profile databases, respectively. The eigenphrase for a bass guitar is a combination of the three profiles of eigenphrase of bass guitar in terms of the onset, interval, and dynamics of a bass guitar. We propose an arrangement method for the bass guitar part by multiplying each principal component vector from the eigenphrase of bass guitar with relative weights. The generated performances from automatic arrangement using the proposed method were confirmed as natural were judged by the authors as natural not artificial.