Recently, Deep Neural Networks (DNN) has been widely used for pattern recognition and classification applications because of its high accuracy. Here in this paper, we propose four different Deep Neural Network (DNN) architectures and comparison is made between these four proposed DNN architectures in terms of accuracy and training time. The proposed DNN models are evaluated for speech recognition application using TIDIGITS corpus. Mel-Frequency Cepstral Coefficients (MFCC) technique is used to extract feature vectors of speech data. It is observed that modified triangular architecture gave the highest accuracy of 99.31 % as compared to other architectures while the triangular architecture gave the least training time of 49.72 sec. Furthermore, results of proposed DNN architecture is compared with the existing Hidden Markov Model based speech recognition and the proposed DNN provide an increased accuracy of 2.33%. © 2018 IEEE.