Enhancement automatic speech recognition by deep neural networks

Abstract

The performance of speech recognition tasks utilizing systems based on deep learning has improved dramatically in recent years by utilizing different deep designs and learning methodologies. A popular way to boosting the number of training data is called Data Augmentation (DA), and research shows that using DA is effective in teaching neural network models how to make invariant predictions. furthermore, EM approaches have piqued machine-learning researchers' attention as a means of improving classifier performance. In this study, have been presented a unique deep neural network speech recognition that employs both EM and DA approaches to improve the system's prediction accuracy. firstly, reveal an approach based on vocal tract length disturbance that already exists and then propose a Feature perturbation is an alternative Data Augmentation approach. in order to make amendment training data sets. This is followed by an integration of the posterior probabilities obtained from several DNN acoustic models trained on diverse datasets. The study's findings reveal that the proposed system's recognition skills have improved.