Modeling processing inaccuracies in human binaural processing of speech using intentional interaural misalignment

Abstract

Human speech intelligibility (SI) in noise improves when target speech and interferers differ in interaural level and time differences, a phenomenon predicted by the equalization-and-cancellation (EC) model. To avoid overestimating binaural benefit, human processing inaccuracies are modeled stochastically using Monte Carlo simulation (MCS) by some models. This approach is computationally expensive and unsuitable for many applications. This study introduces a deterministic misalignment approach to model these inaccuracies in a blind, real-time capable framework. Instead of stochastic sampling, interaural time and level differences are overcompensated by a certain amount to represent average perceptual uncertainty. The misaligned equalization-and-cancellation (EC) stage is combined with a blind back end that estimates intelligibility from phone recognition posteriors, a technique borrowed from automatic speech recognition (ASR) without recognizing the speech itself. The model is evaluated across diverse acoustic conditions, including near- and far-field speech, varying noise directions, non-frontal targets, and interaurally phase-inverted speech in diotic noise. Results show that the misaligned model achieves prediction accuracy comparable to non-blind reference models, without requiring clean speech or noise signals. The approach enables accurate, efficient, and blind prediction of speech reception thresholds (SRTs), making it suitable for continuous monitoring or even real-time implementation in hearing aids and adaptive audio playback systems.