Sound Source Localization Based on Convolutional Neural Network

Abstract

As one of the important ways to transmit and receive information, sound contains a lot of information, and the application scope of sound covers all aspects of life, because the transmission of sound has directionality. In this paper, a room impulse response (RIR) model is established based on the Mono and Multichannel Recording Database (SMARD), and the sound features are extracted using the generalized cross-correlation phase transformation (GCC-PHAT) algorithm. Based on this model, the sound source localization based on convolutional neural network (CNN) is studied. With the help of the microphone array, the system collects randomly generated sound source signals in confined space, and then transmits the collected sound source signals to the CNN model for training. Classify the signal using a well-trained model and finally get the location of the sound source. Experimental results show that the CNN-based sound source localization algorithm has high positioning accuracy under different reverberation conditions and different signal-to-noise ratio environments.