Research On Critical Technologies Of Speech Recognition In Vehicular Noise Environment

The speech recognition technology has been widely applied in multiple areas,but its further development is limited by the poor recognition rate under specific background noise.The project starting from the front-end processing of speech recognition to solve the the problem of poor speech recognition rate of mainstream algorithms in military vehicle-borne noise environment.Improvement idea is proposed based on the traditional algorithm to reduce noise interference,restore the original speech purity,extract the comprehensive features of the signal,and ultimately increase the recognition accuracy of the whole system.This paper consists of following main parts:An modified method for endpoint detection combining ECR with improved spectral subtraction is presented.The algorithm aims at the characteristics of unpitched sound in a specific vehicle environment by first using a split-band spectral reduction for noise reduction,and then performing double-threshold endpoint detection by ECR value obtained from combining the logarithmic frequency domain energy and the auto-correlated clip angle cosine.It has been experimentally confirmed that there is better detection performance in military vehicle environment for the improved endpoint detection algorithm with a combination of time-frequency domain than using a conventional time-domain or frequency-domain algorithm.An modified feature extraction method associating time with frequency domain is presented.The algorithm firstly proposes to combine Mel-Frequency Cepstrum Coefficients(MFCC)with Inverted Mel-Frequency Cepstrum Coefficients(IMFCC),regarding to the shortcomings of traditional Mel frequency cepstrum coefficients in the middle and high frequency bands with poor ability to extract speech signal features.Secondly,the first-order difference forms of MFCC and IMFCC parameters are further extracted considering that human is more sensitive to dynamic speech signals.Then the above parameters are fused and the best dimensional features are selected by introducing the Fisher’s criterion to form hybrid feature parameters in the frequency domain.Thirdly,on account of the superiority of figurativeness and correctness of the time-domain analysis,three-dimensional time-domain features are utilized in combination with the above-mentioned hybrid feature parameters in the frequency domain to finally form improved feature parameters of time-frequency domain combination.It is experimentally verified that the improved combined time-frequency domain feature parameters have better recognition effect in military vehicle environment than the traditional MFCC parameters.On the basis of HMM model,the emulation demonstration interface of voice recognition system in military vehicle environment is created using MATLAB,and the main aspects of the system and the overall process of speech recognition are demonstrated through the platform.And then the improved endpoint detection and feature extraction algorithms are introduced into the system and compared with the system before the improvement.Through a huge quantity of the testing results,it is indicated that the modified approach has significantly raised the recognition rate and verified the practicality and validity in the military vehicle environment.