Complex versus real-valued neural networks for audio source localisation using simulated and measured datasets

Abstract

Complex-valued neural networks can accept complex-valued data as an input and present an alternative to their real-valued counterparts. This can be advantageous for various audio signal processing applications such as for audio source localisation utilising microphone arrays. This paper builds on previous work aimed at comparing the performance of complex and real valued neural networks under equal operating conditions. Furthermore, this work investigates the performance of both types of networks in a 3D source localisation task. In this work an evaluation is made of the performance of networks that are trained using simulated microphone signals but which are then applied to the outputs of real microphone signals. This has advantages due to the simplicity in creating large datasets for the training phase. Both networks are compared to MUSIC, a common classical localisation technique. Results show that both network types can learn from simulated data to localize measured data, although their performance depends on the features with which the networks are trained

More information

Identifiers

ORCID for Vlad Stefan Paul: orcid.org/0000-0002-5562-6102

ORCID for Jacob Hollebon: orcid.org/0000-0002-4119-4070

ORCID for Philip Nelson: orcid.org/0000-0002-9563-3235

Catalogue record

Export record