ANALYSIS OF COMPLEX-VALUED NEURAL NETWORKS FOR AUDIO SOURCE LOCALISATION

Abstract

An increasing number of studies recently have dealt with novel methods for locating the source of a wide range of acoustic events. Applications such as teleconferencing, human-robot interaction, source separation or speech recognition can make use of the Direction of Arrival (DoA) of a sound source to improve their results. Since most of the newer localisation methods proposed recently make use of neural networks in their estimation of source position, the work will focus on comparing the use of complex-valued neural networks with real-valued networks for localising sound source in different scenarios. The data used for the comparison will be simulated using a number of geometrical microphone arrangements with the acoustic sources placed in the far-field of the microphone arrays. The simulated data used in combination with the chosen range of microphone arrangements will be used to investigate the localisation limits of the chosen algorithms. A particular objective of the work will be to evaluate any potential advantages in using complex-valued neural networks rather than real valued networks.