The performance of Automatic Speech Recognition (ASR) depends on its capability to identify the test patterns best-matched with training patterns in various classes. This matching is highly dependent upon the individual feature extraction technique or combination thereof. Certain advanced feature extraction techniques such as GFCC, BFCC have been reported in the literature (with associated additional problems of accepted bandwidth and optimal number of features) in addition to the commonly used ones such as Mel Frequency Cepstral Coefficient (MFCC) and Perceptual Linear Prediction (PLP) coefficient. MFCC is more suitable for clean environments while PLP performs better when there lies a significant mismatch between training and testing phase. Therefore, this paper proposes a minimalistic approach involving hybrid features (i.e., MFCC+PLP) to overcome shortcomings of each constituent, such as sensitivity to background noise on one hand, and avoid complexity in extracting advanced features, such as GFCC and BFCC etc. on the other hand. These hybrid features can provide favourable or comparable results as compared to those obtained using advanced features in both clean and noisy environments. The other problem of optimizing the number of filter banks for a specified bandwidth is proposed to be accomplished using an evolutionary technique like DE (Differential Evolution) to enable suitable comparisons with the existing literature. Additionally, an advanced classifier viz. Deep Neural Networks (DNN) is used as compared to ones that are more conventional such as Hidden Markov Model (HMM) used in the literature for further improvisation.