Direction Estimation of Instrumental Sound Sources Using Regression Analysis by Convolutional Neural Network

There has been much research on estimating noise and speech source direction, but there have not been many studies on estimating the source direction of instrumental sound sources. In this study, we considered the source direction estimation of a single instrumental sound. Direction estimation of sound sources by the multiple signal classification (MUSIC) method often causes large estimation errors. Then, we propose a technique for estimating the direction of musical instrument sound sources by applying regression analysis using a convolutional neural network (CNN), a type of neural network. We calculated the MUSIC spectrum obtained using MUSIC that uses the fundamental and harmonic components, which have relatively large amplitudes, and we estimated the direction of the sound source using the CNN with these components as input. We achieved this by focusing on the overtone structure of the instrumental sound source. This study demonstrated the effectiveness of this method using simulations in a monaural environment.