In this paper, a method based on deep learning is applied to improve acoustic model for English Automatic Speech Recognition (ASR) system using two main approaches of deep neural network (Hybrid and bottleneck feature). Deep neural networks systems are able to achieve significant improvements over a number of last year system. The experiments are carried out on the dataset containing speeches on Technology, Entertainment, and Design (TED) Talks. The results show that applying Deep neural network decrease the relative error rate by 33% compared to the MFCC baseline system.