Chinese Speech Recognition Technology And Its Application In Speech Separation

With the rapid development of computer technology and artificial intelligence,the demand for information interaction between human and machine is increasing.Speech recognition has become an important application technology in the field of information technology.In recent years,although the performance of speech recognition system has been improved continuously,there are environmental noise and speech interference in most application scenarios of speech recognition system,which will seriously affect the performance of speech recognition system.Based on the research goal of "speech separation recognition technology for complex environment",this thesis aims to optimize speech separation technology through continuous speech recognition and improve the ability of speech information acquisition in complex environment.The main work and innovation of this thesis are as follows(1)The realization of Chinese continuous speech recognition.In order to obtain the characteristic parameters of speech signal required for speech separation,the traditional speech recognition process is divided into speech segmentation and speech recognition.In this thesis,the characteristics of speech signal in time domain,frequency domain and cepstrum domain are analyzed,and the voiced segmentation based on pitch cycle trajectory is studied.By combining endpoint detection technology and frequency band energy method,the voiced segmentation is realized.Then,according to the characteristics of Chinese syllables,the combination of voiced and voiced sounds is carried out,and finally the Chinese continuous speech segmentation algorithm is realized.This algorithm is not dependent on the model,not only complete the definition of syllable boundary,but also obtain the priori information needed for speech separation.Compared with similar algorithms,this algorithm has better performance.On this basis,the acoustic model based on VGG-16 and the language model based on N-element grammar are implemented in this thesis.Finally,the Chinese continuous speech recognition is realized.(2)Implementation of CASA speech separation.In this thesis,continuous speech segmentation algorithm is used to obtain pitch cycle trajectory and syllable boundary.By calculating fundamental frequency and harmonics,CASA-based single channel speech separation is realized.This method does not depend on the model and does not need to obtain the priori information of speech and noise.Theoretically,it can be applied to any noisy environment.Subsequently,the performance of various signal-noise separation methods is compared in this thesis.The results show that CASA speech separation has a good noise removal effect in non-stationary noise environment,and it is confirmed that the voice reconstruction algorithm based on speech recognition has a very good performance improvement for CASA speech separation.(3)This thesis analyzes the difficulties and optimization scheme of CASA speech separation technology,and discusses the shortcomings of voice reconstruction algorithm based on speech recognition.In order to realize the sound reconstruction of separated speech,the acoustic model of separated speech was trained.The experiment shows that the acoustic model trained with separated speech has a higher recognition accuracy in the medium and low SNR environment.Then,the acoustic model and the language model were used to optimize the sound reconstruction algorithm.Experiments show that the optimized sound reconstruction algorithm has better noise robustness.Implementation of Chinese speech signal processing system.By integrating the techniques of continuous speech segmentation,speech recognition and speech separation,a Chinese speech signal processing system which can run independently is realized.The test results show that the system runs stably and the output results are ideal.(4)Design the Implementation of Chinese Speech Recognition and Speech Separation System.Integrating the mature research results in this thesis,a system which can run independently is realized.The system can complete speech segmentation,speech recognition,speech separation and other work.