Abstarct: Considering the extensive applications of speech recognition systems in different environmental conditions, strengthening and enhancing the performance of these systems against different types of noise is essential. One way to improve the performance of these systems is to strengthen the feature extraction methods from the speech signal. Most of the existing feature extraction methods in speech processing consider magnitude-baxsed features and ignore the use of phase-baxsed features of speech for various reasons. The main purpose of this thesis is to use a new phase-baxsed feature extraction method to improve the performance of a speech recognition system in noisy conditions. In this regard, we present a novel feature extraction method by changing the structure of the Modified Power-Normalized Cepstral Coefficients (MPNCC) and adding the Modified Group Delay (MGD) to it. The performance of the proposed method is measured by utilizing a Multilxayer Perceptron (MLP) neural network in MATLAB software and using the TIIMT databaxse in order to do word recognition tasks in clean and noisy conditions. In order to compare the performance of the proposed method, two commonly used features in speech recognition, namely Mel-Frequency Cepstral Coefficients (MFCC) and MPNCC, are used under the same conditions. The experimental results demonstrate better recognition accuracy of the proposed method compared to the MPNCC method and is comparable to the MFCC method in noise-free conditions. Under noisy conditions with different SNRs, the accuracy of the proposed method was better than MFCC and MPNCC. For instance, the accuracy of the proposed method in white Gaussian noise at -5dB has improved by 14.8% and 9.7% compared to MFCC and MPNCC, respectively.