Auditory models comparison for horizontal localization of concurrent speakers in adverse acoustic scenarios

This paper aims at comparing and reproducing the predictions of two public available computational auditory models for speaker localization in different simulated environments. The direction-of-arrival (DOA) of sound sources in the horizontal plane can be extracted by using binaural spatial cues from room and user acoustics. Since our predictions consider the specificity of both models at the level of peripheral processing, the proposed solution for DOA extraction also provides a common multi-conditional training for the Gaussian Mixture Model (GMM) approach. A set of acoustic simulations of adverse conditions (i.e. multi speakers or high reverberant scenarios) supports the evaluation phase on robustness of the synthetic auditory process. Our analysis reproduces two case studies from the scientific literature in order to investigate the reliability of localization predictions in the frontal horizontal plane. Finally, a newly defined acoustic scenario allows to identify differences between auditory models outcome in the entire horizontal plane. The results show a good agreement with previous literature and our machine learning approach emphasizes peculiarities of each approach for auditory peripheral processing.