Sound Source Separation Using Spectrogram Analysis by Neural Networks

The performance of neural networks has been dramatically improved since the method called “deep leaning” was developed around 2006[1][2]. Mainly, neural networks have been used for classification problems such as visual pattern recognition and speech recognition. However, there are not so many studies of sound source separation using neural networks. To apply neural networks to separation problems, separation problems require to be transformed into classification problems. To realize it, we referred to spectrogram analysis by specialists. Specialists can separate each source signal from the spectrogram of mixed signals by focusing on each local area of the spectrogram. In this study, we developed a novel method for sound source separation using spectrogram analysis by neural networks. As a result of the simulation, we successfully separated male and female voices from their mixed sound. The proposed method is superior to conventional methods in separation problems with sound reflection on walls and convolutional mixture which includes the difference of traveling time from a sound source to microphones because the method does not require to identify the mixture process in space.