A Speech Graph Signal Feature Detection And De-noising System Based On Forgotten Factors And Graph Sub-band Analysis

With the advent of the 5G era,big data,Internet of things,artificial intelligence and other industries are growing vigorously,and the society is also pushed towards the beautiful vision of intelligence and digitization.Because massive data has gradually become an important data scene faced by all walks of life,it has become a significant research topic to realize the efficient processing of high-dimensional irregular data.In recent years,the new Graph signal processing(GSP)technology provides a new way to solve the difficulties encountered in current signal processing.GSP technology constructs the graph topology model according to the complex scene of generating data,and maps the high-dimensional irregular data to the graph for effective processing according to the correlation between nodes.At present,GSP technology has produced a set of relatively solid basic theories,and has achieved many research results and application results in many cases such as wireless sensor networks,machine learning,social networks and so on.Similar to the classical digital signal processing(DSP),the theoretical framework of GSP technology also includes many key theories such as time-frequency transform,convolution,sampling and filtering.However,different from DSP,GSP technology makes use of the relationships between nodes and has its unique adaptability to different kinds of signal processing.Although GSP technology has some successful application results,these applications are mainly implemented in the scenarios where the prior topologies of the graphs are known,and there is little research on the application scenario where the prior topology of the graph is unknown.Speech signal is one of the most direct and efficient ways of information communication,which has very important research value.At the same time,the traditional DSP technology has made considerable achievements in the field of speech signal processing,and GSP technology has the technical characteristics that DSP technology does not have,so it is very feasible to combine the two.However,as a time-series signal,the graph topology of speech signal is difficult to obtain directly,which is worthy of in-depth study.Therefore,this thesis focuses on the graph construction of speech signal,further applies GSP technology to speech endpoint detection and speech de-noising,and constructs a speech de-noising system based on GSP technology.The main work and innovation of this thesis are as follows:(1)A new speech graph signal structure based on the forgotten factor is proposed.Firstly,according to the correlation characteristics between speech signal nodes,that is,the correlation between adjacent samples is strong,the longer the time interval,the lower the correlation of samples.In this thesis,a speech forgotten graph signal topology is designed by using the forgotten factor and the forgotten threshold.Then,based on the forgotten graph topology,the graph adjacency matrix of speech graph signal is obtained,and the basis function of speech graph frequency domain based on this matrix is further obtained.By analyzing the spectral distribution characteristics of speech signal in the frequency domain of the new speech graph,and combining this new characteristic with the concept of spectral entropy of classical speech signal processing,a graph adaptive band-partitioning spectral entropy(GABSE)algorithm based on the forgotten factor is proposed to detect the endpoint of noisy speech signal in noisy environment.The results of experiments show that,compared with the GABSE algorithm and robust endpoint detection algorithm in classical DSP,the graph adaptive band-partitioning spectral entropy algorithm proposed in this thesis shows better performance in the detection accuracy of speech endpoints in the environment of stationary noise and non-stationary noise with different signal-to-noise ratio.(2)Considering that the speech endpoint detection is often combined with the speech de-noising algorithm to obtain better de-noising performance,a graph adaptive spectral subtraction based on non-uniform graph sub-bands(GASS-NGS)is proposed in this thesis.In view of the non-uniform distribution of speech graph frequency components in the new graph frequency domain,a non-uniform speech graph sub-band division method is proposed in this thesis,which divides the graph frequency into 13 non-uniform speech graph sub-bands with narrow bandwidth in low frequency,wide bandwidth in high frequency.On the basis of dividing non-uniform graph sub-bands,this thesis improves the idea of graph spectral subtraction in the paper,and proposes a GASS-NGS algorithm.The results of experiments show that,the GASS-NGS algorithm proposed in this thesis shows better performance in Objective SNR and subjective PESQ compared with classical sub-band spectral subtraction,Wiener filter method and conventional graph spectrum subtraction.Finally,a speech de-noising system based on GSP technology is designed by combining the GABSE algorithm and the GASS-NGS algorithm.