Sound Source Localization Based on von-Mises-Bernoulli Deep Neural Network
Kazuhiro Nakadai, Shungo Masaki, Ryosuke Kojima, Osamu Sugiyama, Katsutoshi Itoyama, Kenji Nishida · 2020
This paper addresses the learning of periodic information such as the phase for deep neural networks (DNN). To solve this problem, we propose a novel activation function based on the von Mises distribution used in directional statistics, and we construct von-Mises-Bernoulli DNN by replacing the activation function in the first hidden layer of the conventional Bernoulli-Bernoulli DNN with the proposed activation function. We theoretically validate that this simple change can handle periodic information by using restricted Boltzmann machines. In addition, we practically show that the constructed von-Mises-Bernoulli DNN can learn periodic phase information through numerical simulation for sound source localization. Experimental results showed that sound source localization accuracy was almost 100% for 5° resolution, and 60% for 1° resolution under noise condition of 20 dB signal-to-noise ratio.