Deep network source localization and the influence of sensor geometry

Learning-based localization approaches cast the acoustic speaker localization problem as a machine learning task where a classifier is trained on example data of acoustic feature vectors in order to predict likelihood of speech presence as a spatio-temporal distribution. We investigate the impact that fundamental acoustic parameters of the auditory scene (e.g. SNR, acoustic scene complexity, sensor geometry) exert on the ability to faithfully extract spatio-temporal activity maps for concurrent speakers. Our results indicate that to some degree shortcomings in the acoustic conditions can be compensated by increased complexity in the applied classification techniques. To this end, we systematically investigate localization performance for a set of deep neural network localizers of varying complexity, and for six different sensor configurations in a bilateral hearing aid setup. Deep networks result in improved performance compared to linear localizers, and their performance benefits more from an increase in the number of sensor channels. In specific configurations, deep networks with a smaller number of microphones perform better than a linear baseline network with a larger number of microphones. Thus, location-specific information in source-interference scenarios appears to be encoded non-linearly in the soundfield, requiring non-linear approaches for optimal decoding.