Research On Speech Enhancement Technology Based On Graph Signal Processing

The resurgence of artificial intelligence and the vigorous development of the Internet of Things have greatly promoted the informatization process of society,which has improved people's life quality a lot.However,rapid expansion of large-scale irregular data caused by big data era also puts forward higher requirements on signal processing technology in the meanwhile,which promots the development of a new signal processing method: graph signal processing(GSP).GSP processes data indexed by graphs,which are built by exploiting and utilizing information such as internal correlations of data.It has initially formed a basic theoretical system similar to the classic digital signal processing(DSP)technology.GSP has shown its high prospects in the field of signal processing,and has been widely used in wireless sensor networks(WSNs),image processing,machine learning and many other areas.The main difference between GSP and DSP is that GSP takes into account the edges and weights among samples,which is also the significant advantage of GSP over DSP.In this way,GSP fully considers the relationship among various vertices,thus obtains an improvement in processing performance.The research direction of GSP covers many aspects such as the construction of graphs,sampling,filtering,reconstruction,etc.However,most of this work is carried out under a priori assumption that the graph of signal is known,which is also a key reason why GSP is rarely adopted to process signals without obvious graph structures such as speech signals.However,time series like speech play a significant role in people's daily life as one of the important information sources,so the study of its processing technology is undoubtedly of great significance.Moreover,there exist general correlations among the samples of speech signals,which are ignored when describing signals with traditional DSP methods,and this also means that GSP is of great prospects in speech signal processing.Therefore,this thesis mainly focuses on the construction of appropriate graphs for speech signals and the application of GSP in speech enhancement algorithms.The main work and innovations of this thesis are as follows:(1)An iterative spectrum subtraction method based on k-joint shift operator is proposed to suppress noise in noisy speech.The prerequisite and key to leverage GSP on processing speech signals is to determine an appropriate graph so as to map the speech signals into the graph domain.Considering the correlations among adjacent vertices of speech signal,this thesis proposes a joint k-shift operator and views it as the djacency matrix to map speech from time domain to graph domain.Based on the further analysis of the graph spectrum of both speech and noise signals in the corresponding graph frequency domain,an iterative spectral subtraction algorithm is then proposed.The experimental results show that the basic spectral subtraction without iteration outperforms the traditional spectral subtraction in DSP in both signal to noise ratio(SNR)and the perceptual evaluation of speech quality(PESQ).Moreover,with the introduction of the iterative mechanism,the proposed IGSS achieves a better noise suppression effect over the GSS,and has better performance than the basic iterative spectral subtraction methods in DSP.(2)This thesis further studies the Wiener filter in GSP,which is one of the most classic speech enhancement algorithms in DSP,and proposes a graph Wiener filter based on the graph Laplacian operator.Graph is a key element that reflects the relationship among vertices.In order to make full use of the characteristics of speech signals,this thesis designs a dynamic graph under the step constraint to obtain the construction of graph speech signal.By leveraging the graph Laplacian operator,this thesis further analyzes the different distributions of the speech and noise signals in the corresponding graph frequency domain,and then designes a graph Wiener filter with the minimum mean square error(MMSE)to be the optimal criterion.The experimental results show that the proposed algorithm can achieve better noise suppression effects than the classical Wiener filter in both SNR and PESQ.Furthermore,compared with the graph Wiener filter proposed in GSP,our proposed algorithm shows more advantageous for the different SNR conditions,and it applies to a wider range of noise scenarios.