Binaural Speech and Distribution Learning for Predicting Auditory Distance Perception in Audio Augmented Reality

Abstract

This article presents the first experiments aimed at predicting the physical distance that listeners attribute to speakers rendered in Audio Augmented Reality - virtual immersive environments that engage hearing but not other senses. This problem is significant because physical distance between individuals is more than geometric information; it is a factor that both shapes and is shaped by the social relationship between speaker and listener. The primary challenge is that, in the absence of visual cues, people often struggle to localize sound sources accurately and, consequently, to determine the position of a speaker they cannot see. For this reason, the experiments adopt a distribution learning approach, predicting not a single position but a probability density function over the space of possible positions. This approach better reflects the variability in attributed positions observed across multiple listeners for a given speech stimulus. The experiments were conducted using a publicly available corpus of 360 spatially rendered speech stimuli from 120 speakers (60 women and 60 men). The results show that the perceived distance can be predicted with correlation up to 0.79 and Mean Absolute Error as low as 41 cm.