Categorization of Speech Frequency Following Responses to English Vowels

This paper investigates the categorization of the speech Frequency Following Response (sFFR) evoked by the five English vowels using machine learning models. As part of the control system of the brain-controlled hearing aid, the models based on Convolution Neural Networks (CNNs) are used to extract spectrum features from sFFR signals to classify vowels. The highest average accuracy reaches 80.00% for CNN with loose recurrent connections. In addition, we tested the performance of models when there is different additive noise to the input sFFR signals and found that the loose recurrent connections help the CNN model to obtain stronger robustness. We also use motif topology to analyze the content features of the lateral connection network after training. At last, we compared our work with other research. The results of this work show the potential of the proposed network models for the categorization of sFFR signals applied to brain-controlled hearing aids, especially in the presence of noise.