Deep Neural Network Algorithms for Speech Recognition in Interactive Voice Response System

In modern world speech recognition is one of the most popular and developing technology. This paper aims to increase the efficiency of speech recognition via using deep learning technology. During work on this project research there was done the analytical review of the literature of history and classification of speech recognition, as well as about modern approaches and models. Moreover, there was analyzed such models as Hidden Markov models (HMM), Gaussian mixture models (GMM), long shot-memory model (LSTM), deep neural networks (DNN) and hybrid HMM with DNN. Within this work was carried out analytical review of various acoustic models, which was trained by GMM and DNN. In addition to was performed comparative analysis of modern toolkit and experimental research based on Voxforge receipt and Kaldi library. Besides, during the research results was compared according to gender and age feature. The result of the research shows that Kaldi library is more accurate then Google toolkit. Furthermore, models based on DNN are more effective in several times.