Research On Key Technologies Of Speech Recognition In Tank Noise Environment

With the continuous rise of artificial intelligence technology,speech recognition technology as the main communication method in human-computer interaction has gradually attracted people's attention.Due to the short development time of speech recognition technology and the immature part of the technology,it leads to various problems in the specific environment of speech recognition system.In order to solve the above problems,this thesis focuses on the key research of speech recognition system in tank environment,elaborates the basic principles of core technologies including speech preprocessing,endpoint detection and feature extraction.Then some improvement ideas and methods were put forward for some key technologies of speech recognition system.A new detection method combining multitaper spectral estimation and improved energy-entropy ratio based on EMD is proposed.This method firstly uses multitaper spectral estimation on the sampled signal to reduce noise before endpoint detection.Then a new speech parameter is obtained by combining Teager energy of EMD and short-term spectral entropy,which is EMD improved energy-entropy ratio.Finally,the parameters of the improved energy-entropy ratio are used to detect the endpoints of the noisy speech using a dual-threshold algorithm.It's proved that the algorithm has an ideal detection effect in a tank environment.It is proposed that a hybrid feature parameter MGFCC mixes the characteristics of MFCC and GFCC.Firstly,the Teager energy of the sampled signal replaces the traditional time-domain signal to reduce the influence of noise in the speech recognition process.Then,for the shortcomings of MFCC,the hybrid feature parameters of MFCC,MIDMFCC and IMFCC are combined with Fisher criterion to obtain the feature parameter MMFFCC,which can make up for the shortcomings of MFCC in extracting less voice information in the middle and high frequency bands.Next,the Fisher criterion is used to perform feature selection on the GFCC and its first-order dynamic parameters to obtain the hybrid parameter GGFCC,which makes up for the shortcomings of MFCC that cannot extract effective features of speech in a complex environment.Finally,GGFCC and MMFCC are mixed to obtain a new feature parameter MGFCC.After a large number of experiments,it has been proved that MGFCC has better robustness and recognition effect than traditional MFCC and GFCC in tank environment.At the end of this thesis,the definition of HMM and the problems to be solved in the realization of HMM model in speech recognition system were briefly explained.Based on the Hidden Markov Model,a complete tank speech recognition system is built by using MATLAB software,shows the operation interface in the process of speech preprocessing and training recognition.Comparing the improved algorithm with the traditional algorithm,experiments show that the improved algorithm can significantly improve the recognition rate of the voice system in a tank environment.