Training Methods for Deep Neural Network-Based Acoustic Models in Speech Recognition

Nowadays, speech recognition technology is built on Deep Neural Networks. These networks represents the latest direction of machine learning. They are based on the theory of artificial neural networks, which have been used for decades. However, unlike traditional Neural Networks, all deep networks contain many processing layers, which allow the hierarchical processing of the input data. While the concept of deep networks is not totally new, their efficient training required several new achievements. These new networks managed to completely replace the Gaussian Mixture Models in the state-of-the-art speech recognition systems. In this study, we decided to focus on Deep Neural Network-based recognition systems. First, we compared the performance of several new training algorithms with each other, in order to determine the best one for later use. Then, we turned my attention to the algorithms that the new speech recognition systems have inherited from the previous Gaussian Mixture Model-based approaches, as the algorithms might not be optimal for Deep Neural Networks. we proposed new algorithms for obtaining the initial alignment of the frame-level state labels and the creation of context-dependent states, and found that they are better suited for the new acoustic models. Lastly, we also experimented with a data re-sampling method to improve the accuracy of the models.