Abstarct: Environmental noise is an inevitability for everyday life which in turn causes speech signal deterioration and speech recognition accuracy reduction that is a basic topic of speech processing. It's necessary to reinforce these systems against various types of noise baxsed on their widespread application in different environments. The feature extraction process is the first and foremost component of these systems in which area numerous noise removal algorithms have been introduced in recent years. In this study, three new feature extraction methods, BFMHECC, HEBFCC1, HEBFCC2 baxsed on the Mean Hilbert Envelope coefficients and Bark filter bank cepstral coefficients which is inspired from human auditory system, are presented. Efficiency of feature coefficients extracted by the presented innovative methods is compared with other widely used feature extraction methods, implementing the multi-laxyer perceptron neural network method for speech recognition using the TI_DIGIT databaxse on noisy and clear speech signals. Comparison of results obtained implementing these methods illustrate a more or less better performance by the proposed methods in noisy and clear speech signal compared to conventional approaches.among the three proposed methods, the BFMHECC approach yields better performance ratings compared to the other two, under various noises with SNR values of -5, 0, 5 and 10 baxsed on accuracy evaluation techniques and EER.