In P300 based BCI systems, eliciting ERP using the oddball stimulation will conceal the original P300 components in EEG signal. Therefore, it requires accurate detection of P300 components to precisely recognize the characters. For that purpose, conventional machine learning and its ensemble variant have been implemented in the existing Devanagari Script (DS) based P300 speller to classify P300 components. However, the performance of the conventional machine learning techniques had degraded while detection of P300 components in reduced number of trials due to its the inability to poorly identifies and handles non-linearities of EEG signals. Hence, it increased the trade-off between accuracy and spelling time in existing DS based P300 speller. Unlikely, deep learning approaches have provided state of art performance for classification and pattern recognition in other EEG signal-based applications. Therefore, in this work, we have implemented two proven deep learning algorithms i.e. deep convolution neural network (DCNN) and stacked autoencoder (SAE), customized and fine-tuned for classification of P300 and non-P300 components in DS based P300 speller. Also, we have implemented a novel double batch training approach to handle the computational burden. Additionally, a leaky ReLU activation function is used in DCNN to overcome dying ReLU problem. The experiments have implemented on self-generated dataset of 20 Devanagari words with 79 characters acquired from 10 subjects using 16 channel ActiCAP Xpress EEG recorder. The accuracies of 84.85% & 88.22% are obtained for SAE & DCNN respectively with just three trials. Moreover, the time to spell a character is reduced from 16 secs to 9.6 sec with improvement in information transfer rate (ITR) from 10.07 to 20.6 BPM using the proposed DCNN.