The noise of cockpit voice information has high loudness, numerous types and wide frequency range, which has serious influence on performance of cockpit voice recognition.To solve the problem, an adaptive filter based on Least Mean Square algorithm was used for noise reduction, which could achieve the best noise reduction effect by adjusting the order and the step length of the filter.After that, the cockpit voice was pre-emphasized, framed, windowed and conducted Fourier transform, followed by extracting Mel-Frequency Cepstrum Coefficients (MFCC) and first-order differential cepstrum parameters as feature vectors.Finally, a Support Vector Machine (SVM) was designed for training and identification.The problem that the performance of cockpit voice recognition is poor under a low SNR was solved.The simulation results show that this method is obviously superior to the wavelet packet de-noising, and the recognition accuracy rate reaches 96.9231%.