Handwritten word recognition from document images using deep learning is an active research area in the field of Document Image Analysis and Recognition. In the present era of Big data, since more and more documents are being generated and archived in the compressed form to provide better storage and transmission efficiencies, the problem of word recognition in the respective compressed domain without decompression becomes very challenging. The traditional methods employ decompression and then apply learning algorithms over them, therefore, novel algorithms are to be designed in order to apply learning techniques directly in the compressed representations/domains. In this direction, this research paper proposes a novel HWRCNet model for handwritten word recognition directly in the compressed domain specifically focusing on JPEG format. The proposed model combines the Convolutional Neural Network (CNN) and Bi-Directional Long Short Term Memory (BiLSTM) based Recurrent Neural Network (RNN). Basically, we train the model using JPEG compressed word images and observe a very appealing performance with 89.05% word recognition accuracy and 13.37% character error rate.