Performance Evaluation of a Simple Deep Neural Network System for Auditory Attention Decoding

Recent years have seen an increase in the variety of methods used to perform Auditory Attention Decoding (AAD). Current high performing methods for auditory attention decoding rely on large training sets and lack comparable standards with one another largely due to the variability in the training data used. Simple standards between these models could help researchers better interpret performance and direct the progression of work to a model that performs the best. Here the performance of a Deep Neural Network (DNN) architecture for AAD proposed by (Cicarelli et al.,2019) is evaluated on a new, smaller set of training data collected in (Fuglsang et al.,2017). The network is shown to successfully achieve learning behavior when presented with the reduction of training data. Limiting the number of listeners used for training based on the output average loss curve resulted in comparable decoding accuracy. Further metrics show the benefit of an analysis of the relevance of listeners used for training of the network. The consistent performance of the network given the reduction in the provided training data shows how the simple DNN is a robust method for performing AAD. It also allows us to properly compare the performance of the DNN with the linear method in (Fuglsang et al.,2017)