| At present,the application of speech technology in human production and life is becoming more and more extensive.However,with the complexity of the production and living environment,the application of speech technology in production and life is also facing more and more challenges.In the application of speech technology,speech is disturbed by noise in the environment,resulting in reduced speech quality and intelligibility,which further leads to a decrease in speech recognition rate in human-computer interaction.As a front-end module in human-computer interaction,speech enhancement technology plays an important role in ensuring speech quality.For speech enhancement,this thesis mainly uses graph Fourier transform(GFT)as the feature processing method of speech time series signals,and does the following three aspects.First,the graph Fourier transform is combined with non-negative matrix factorization(NMF)to propose a speech enhancement algorithm based on GFT-NMF.Firstly,the speech timing signal is framed and the corresponding graph signal is constructed,and the graph signal will be mapped to the graph frequency domain after the Fourier transform.In the model training stage,the amplitude spectrum of the clean speech graph spectrum is input into the NMF model,and finally the effective decomposition of the amplitude spectrum in the clean speech spectrum by the NMF model is realized.In the speech enhancement stage,the amplitude spectrum of the noisy speech spectrum is used as the model input,and the effective decomposition parameters obtained in the training stage are combined to realize the enhancement of the noisy speech.In this thesis,a speech enhancement algorithm based on short-time Fourier transform-non-negative matrix factorization(STFT-NMF)is used as the baseline for comparative experiments.Experimental results show that the speech enhancement algorithm based on GFT-NMF can obtain better enhanced speech quality than the speech enhancement algorithm based on STFT-NMF under high signal-to-noise ratio.Second,after the GFT-NMF algorithm,this thesis combines the graph Fourier transform with the fully convolutional neural network(FCNN)to propose a speech enhancement algorithm based on GFT-FCNN.In this algorithm,the noisy speech graph spectrum is used as the model input,the clean speech graph spectrum is used as the expected output of the model,and the ratio mask is used as the training target for model training.In the FCNN model,the input layer,encoder,decoder and output layer are included,and the encoder and decoder can be used to mine the graph frequency information in depth to achieve better speech enhancement effect.Experimental results show that the speech enhancement algorithm based on GFT-FCNN shows better performance in speech quality and intelligibility than the STFT-FCNN algorithm.Third,with GFT-NMF algorithm as the background core algorithm module,a speech enhancement system is designed and implemented by MATLAB technology.The system includes three parts: model training module,batch speech enhancement module and single speech enhancement module,which can realize model training in speech enhancement tasks and enhance noisy speech. |
