Model Analysis of Deep Learning for Sound Image Direction Perception Using Explainable Artificial Intelligence
Year: 2026
Type: article
Abstract: In recent years, deep learning models of human auditory perception have been researched for a super smart society. Deep learning has the potential to improve estimation accuracy; however, its internal structure becomes complex and difficult to explain, making improvements in explainability essential in the use of safety‐critical systems. In a previous study, deep learning binaural models of three‐dimensional sound image direction perception, which is an important factor of auditory perception, were constructed using a convolutional neural network in simulation environments that take into account in‐vehicle sound fields, where it is difficult to perceive sound image direction, in‐room sound fields, and free‐field conditions. In this study, the influence of sound source characteristics, the three‐dimensional sound source direction, and the sound fields on the model learned in a similar way was analyzed. The results of statistical analysis of the relationship between the features extracted from the weight distribution by local interpretable model‐agnostic explanations, an explainable artificial intelligence method, and the features of sound image direction perception of humans confirmed that the convolutional neural network obtains the features of sound image direction perception of humans. Moreover, improvements to the model's accuracy and computational cost were investigated by utilizing the model analysis results. © 2026 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.
DOI: 10.1002/tee.70418
Author: Koji Sakamoto
Institution: Osaka Sangyo University
Language: English
Topic: Speech and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7204544265