Speech-Aware Binaural DOA Estimation Utilizing Periodicity and Spatial Features in Convolutional Neural Networks

In recent years, several supervised learning-based approaches have been proposed for estimating the direction of arrival (DOA) of a single talker in noisy and reverberant environments. In the absence of auxiliary information, such as a voice activity detector (VAD), the estimated DOA may be erroneous due to speech pauses or noise dominance. In this paper, we consider a speech-aware DOA estimation system for binaural hearing aids, which does not require a separate VAD. This system utilizes a combination of spatial features with an auditory-inspired periodicity feature called periodicity degree (PD) as input features of a convolutional neural network (CNN). Using speech and non-speech signals during the training, the CNN can capture the harmonic structure encoded in the PD features, thereby distinguishing speech from non-speech portions and simultaneously mapping spatial features to sound source DOA upon speech detection. To investigate the benefit of using PD features for speech-aware DOA estimation, we evaluated the performance of speech-aware systems that utilized either broadband or narrowband feature combinations compared to baseline systems. We propose to use a novel narrowband feature combination consisting of the narrowband cross-power spectrum (CPS) as the spatial feature and a new subband-averaged representation of PD features. The broadband feature combination consisted of the generalized cross-correlation with phase transform (GCC-PHAT) and the broadband PD features. The baseline systems considered in this work consisted of a CNN that exploits only a spatial feature, cascaded with a VAD. Evaluations in reverberant environments with different background noises for both static and dynamic single-talker scenarios demonstrate that incorporating the PD feature in conjunction with any type of spatial feature provides an advantage for binaural DOA estimation in terms of accuracy and angular error.