Sound source localization using complex-valued deep neural networks
Gwantae Kim, David K. Han, Hanseok Ko · 2024
This paper presents a complex-valued deep neural network for sound source localization. Most neural network-based sound source localization approaches use time-frequency domain features. Even though both magnitude and phase play a pivotal role in solving the sound source localization problem, the real-valued features are only used because the neural network structures generally accept real-valued inputs only. In contrast, the complex-valued neural network structures directly receive complex-valued inputs and extract complex-valued hidden features. Therefore, the complex-valued deep neural network, which is proposed in this paper, has the potential to extract rich features for sound source localization. With a series of experiments, the proposed direction of arrival estimation method with a complex-valued deep neural network outperforms the real-valued deep neural network-based method.