Spoken Digit Recognition Using Convolutional Neural Network

Abstract

This paper presents a spoken digit recognition study based on features extracted from Bark spectrogram and classifies it using convolutional neural network (CNN). In this work, training and test databases of the spoken digits were developed from TI20 corpus. 6,506 spoken digits from 16 individuals have been used in the study. A 198 x 50 feature map has been extracted from each spoken digit to be fed as input to the CNN. A 5-layer convolutional filters have been added as layers in the CNN architecture. Adam optimization algorithm is then used to train the CNN. A confusion matrix resulting from classification on the test database has been presented in this paper. It is shown in this paper that the proposed CNN architecture performing on the test database achieved an average accuracy of over 99% for the known spoken digits and wrongly classify as digit up to 3.4% for the unknown spoken words.