Speech recognition is a method where an audio signal is translated into text, words, or commands and also tells how the speech is recognized. Recently, many deep learning models have been adopted for automatic speech recognition and proved more effective than traditional machine learning methods like Artificial Neural Networks(ANN). This work examines the efficient learning architectures of features by different deep neural networks. In this paper, five neural network models, namely, CNN, LSTM, Bi-LSTM, GRU, and CONV-LSTM, for the comparative study. We trained the networks using Audio MNIST dataset for three different iterations and evaluated them based on performance metrics. Experimentally, CNN and Conv-LSTM network model consistently offers the best performance based on MFCC Features.