Binary-Temporal Convolutional Neural Network for Multi-Class Auditory Spatial Attention Detection

Humans have a remarkable ability to focus on one of the sound sources in a multi-speaker environment. Auditory spatial attention detection (ASAD) aims to identify the direction of the speech source a person is attending to based on their brain signals, with potential applications in enhancing hearing aids, improving communication systems, and advancing brain-computer interface (BCI) technologies. Most prior studies formulated the problem as binary classification, however, real-world scenarios are much more complex. Our study explores the feasibility of detecting auditory attention among 10 competing speakers. To address the needs of low-resource computing equipment, we further propose a novel approach using a binary temporal convolutional neural network (B-TCNN) for multi-class ASAD tasks. This study effectively reduces memory consumption and accelerates inference. Experimental results show that the B-TCNN achieves an average accuracy of 93.8% with only 33K parameters in 1-second decision windows for a 10-class ASAD dataset. The proposed network significantly outperforms other competitive models, offering a lightweight and efficient solution for multi-class ASAD tasks.