Research On Speech Enhancement Technology In Driving Environment

As technology continues to evolve,communication between people and machines can already be achieved through voice.However,there are ambient noise and reverberations in the process of passing voice commands to the machine,which affects the machine's acceptance of signals and correct command decisions.This thesis will focus on the effectiveness and pertinence of speech recognition technology in the driving environment.The technology includes preprocessing,speech noise reduction technology and feature parameter extraction.The specific research and improvement results are as follows:An improved algorithm based on spectral parameter value and spectral subtraction combined with masking method is proposed.The method does not achieve the optimal parameters of the traditional spectral subtraction formula,and uses the spectral parameters of the speech to calculate the parameters suitable for the ambient speech,and then combines with the masking method to finally obtain the noise-reduced speech.Experiments show that the improved algorithm has good noise reduction effect in the driving environment and can effectively remove the "music noise" generated by the traditional spectral subtraction.An improved feature extraction method based on multi-layer wavelet packet decomposition is proposed.The method firstly decomposes the wavelet packet into 6layers,and decomposes 1-2 layers on the basis of the original.The frequencies of the unvoiced and voiced sounds are different,and the frequency distribution after the wavelet packet change can be accurately arranged.Then calculate the average energy of each frequency band,combine the MFCC technology to find the cepstrum,and finally apply the Fisher ratio to select the best feature parameters.Experiments show that the improved algorithm has a large improvement in the recognition rate.In the driving environment,the detection and recognition rate can reach more than 90%.The thesis introduces the basic theory of hidden Markov and the three problems and solutions encountered in modeling.The HMM-based speech recognition system was built on Matlab and compared with the previous improved algorithms.After a large amount of experimental data analysis and comparison,it is concluded that the improved system has a greater improvement in recognition rate than the original identification system.