Research On Key Technologies Of Speech Recognition In Airborne Noise Environment

Because the speech signal is mixed with noise and other interference signals,the recognition system can not accurately receive and judge instructions.This thesis mainly studies how to improve the accuracy and robustness of speech recognition in airborne noise environment,and improves the key technologies in speech recognition,such as noise reduction,endpoint detection,feature extraction and so on.Aiming at the problem of noise interference in speech recognition,an algorithm combining EMD decomposition and LMS noise reduction is designed.Firstly,the empirical mode decomposition(EMD)method is studied and analyzed,and then the traditional least mean square(LMS)algorithm is improved.The algorithm makes use of the characteristics of EMD multi-dimensional decomposition and LMS adaptability,which not only overcomes the defect of poor LMS denoising effect,but also offsets the residual noise after LMS denoising.Finally,the simulation results show that compared with the traditional noise reduction algorithm,the signal-to-noise ratio of the improved LMS algorithm can increase 2db～4db more than the traditional noise reduction technology in different types of noise environments.In order to solve the problem of misjudgment and missed judgment of endpoint detection in speech recognition at low signal-to-noise ratio,a difference to zero ratio detection method is designed.Firstly,the wavelet packet balk subband variance endpoint detection method is studied and analyzed,and then the short-time zero crossing rate is introduced to combine the two characteristic parameters,namely short time difference zero.Finally,combined with the above improved noise reduction algorithm,the simulation results show that the accuracy of the improved algorithm is improved by 10%～30% compared with the traditional algorithm in different types of low SNR noise environment.Aiming at the weak generalization ability of feature extraction in speech recognition,a feature extraction method based on EMD is designed.Firstly,the linear prediction cepstral coefficients(LPCC)and Mel frequency cepstral coefficients(MFCC)are studied and analyzed.Then,according to the characteristics that the noisy information components after EMD decomposition are mainly concentrated in the first few stages,the noisy components in the speech signal after noise reduction detection are unmarked,and the remaining information components are processed to improve MFCC.Finally,the simulation results show that the recognition rate of the improved MFCC feature parameters in airborne noise environment with different signal-to-noise ratio is 10%～30% higher than that of the traditional feature parameters.A speech recognition platform is based on Hidden Markov model(HMM).Firstly,the airborne speech recognition system is built on MATLAB to display the speech recording,preprocessing,noise reduction detection,feature extraction,training,recognition and so on in detail.Then the experimental data of the new algorithm are counted and compared with the recognition rate of the traditional algorithm.The feasibility of the new algorithm in airborne noise environment is verified by experiments,which has certain practical value.