Concurrent Vowel Identification Using the Deep Neural Network

Abstract

An alternative deep neural network model was developed to predict the effect of fundamental frequency (F0) difference on the identification of both vowels in concurrent vowel identification experiment. In the current study, the time-varying discharge rates, computed from the auditory-nerve model, to concurrent vowel were the inputs to a ten-layer perceptron. The perceptron was trained, using the gradient descent with momentum and adaptive learning rate backpropagation algorithm, to obtain a similar identification score observed in normal-hearing listeners for 1-semitone. The perceptron was then tested to predict the concurrent vowel scores for the other 5 F0 difference conditions. The proposed perceptron model was successful in qualitatively predicting the concurrent vowel scores across F0 differences, as observed in concurrent vowel data.