Localization of Steady Sound Source and Direction Detection of Moving Sound Source using CNN

This paper proposes a convolutional neural network (CNN) based classification method for broadband direction of arrival (DOA) estimation of steady sound source in noisy conditions and also in reverberation conditions using a uniform linear array (ULA) of microphones. In addition, we also find out the direction of moving sound source (left or right). The input to the CNN is given as the Short-Time Fourier Transform (STFT) coefficients of the phase components obtained from the ULA of microphones. The CNN then learns the features required for training. Here we have used the room impulse response (RIR) of each angle and White Gaussian Noise of different variances to generate training database. Also we add synthesized noise signal to training data set to generate the actual speech signal of that particular room, so that the CNN can classify speech source according to the DOA during the demonstration.