音源信号の伝播時間を考慮したニューラルネットワークによるブラインド音源分離

抄録

Blind source separation is an approach for estimating source signals that uses only the mixed signal information observed at each sensor. Independent component analysis (ICA) is a powerful way to solve this problem if it is applied to instantaneous mixture in which travel time is not considered. If we want to solve the mixture in which travel time of source signals is considered, it has to be dealt in frequency domain. This way makes it more difficult to solve the problem in real time. Our approach deals the signals in time domain by neural networks. Neural networks have ability to predict by giving time series data as input. Therefore, we assume that it can predict the separated source from the time series data of mixed signals. As a result of simulation, we successfully separated the mixture of sine waves and single voices “a” which are spoken by male and female. Additionally, even if there are long time difference in traveling, we separated the mixture of sine waves.