Received: 30-12-2019 / Accepted: 26-09-2020
Speech recognition has been attracting many researchers in the field of artificial intelligence recently. For example, the problem of implementing a program for robots to recognize human speech, thereby robots can understand, learn and talk with human. In this study, 37 students from Vietnam National University of Agriculture were involved to acquire speech data of 29 letters in Vietnamese alphabet. The data were preprocessed to extract featured voice chunks for the classification. We then used the deep Boltzmann machine (DBM) as a deep network with stacked hidden layers. To evaluate the proposed method, we compared the learning performance of DBM to a neural network (NN) with the same network structure configuration. The results showed that DBM performed better with accuracies of 68% on the training dataset and 51% on the test dataset, while the respective figures for NN were 61% of training and 48%.