A Two-Stage CNN with Feature Reduction for Speech-Aware Binaural DOA Estimation
Reza Varzandeh, Simon Doclo, Volker Hohmann · 2023
In recent years, several supervised learning-based approaches have been proposed to estimate the direction of arrival (DOA) of a single talker in noisy and reverberant environments. In this paper, we consider a speech-aware DOA estimation system for binaural hearing aids, which does not require a separate voice activity detector (VAD). We propose the combination of two narrowband features as the input features of a convolutional neural network (CNN), namely the cross-power spectrum as spatial features and narrowband auditory-inspired periodicity features. Prior to the joint processing of both features, we propose to reduce the dimensionality of the narrowband periodicity features using a feature reduction stage based on$1\times 1$convolutions. Simulation results for two reverberant environments with different background noises demonstrate the benefit of the feature reduction stage in terms of DOA estimation accuracy while significantly reducing the number of trainable parameters. In addition, simulation results show that the proposed system outperforms a baseline system consisting of a CNN using only spatial features and a pitch-based VAD.