Shortwave Single Sideband Speech Enhancement Technology Based On Deep Learning

Shortwave communication has the characteristics of long transmission distance,low cost,high reliability,flexible maneuverability,and strong invulnerability.Single sideband speech signal is the main transmission form of shortwave speech communication and is widely used in military and civilian communication fields.Due to the long-distance transmission of shortwave speech signals through the ionosphere,the received speech signals are weak,and at the same time affected by channel fading,background noise and interference,the quality of received shortwave speech is very poor.However,the traditional speech enhancement technology based on statistical characteristics is difficult to accurately estimate the shortwave noise,which leads to the distortion of the enhanced speech signal.The emergence of deep learning provides a new solution for speech enhancement.A large number of speech enhancement network models have emerged one after another,and have achieved superior performance than traditional speech enhancement algorithms.Therefore,thesis mainly studies the short-wave single sideband speech enhancement technology based on deep learning.In thesis,aiming at the problem of speech signal distortion caused by traditional speech enhancement methods,on the basis of the classic noise reduction neural network structure,the gating mechanism is applied to the GCRN network for intelligent noise reduction,and a multi-branch dynamic merge network and CEB＿LSTM network are proposed.Combining the powerful mapping ability of the Convolutional Neural Network and the advantages of the Recurrent Neural Network in processing time series,compared with the GCRN network,the PESQ score under the average signal-to-noise ratio has increased by about 0.18.Compared with the CEB＿STCM network,the PESQ score under the average signal-to-noise ratio has increased Around 0.1,through the verification of simulation data and actually collected shortwave speech signals,CEB＿LSTM network has significantly improved network noise reduction performance and generalization ability.At the same time,thesis aims at the problem of part-band interference in shortwave channels,and generates a CEB＿LSTM anti-jamming network,which can effectively eliminate background noise and part-band interference.Both the precision rate and recall rate exceed 95%,and significantly improves the quality and intelligibility of speech signals.Aiming at the problem that extremely weak speech signals are easily regarded as noise and eliminated in the short-wave speech enhancement technology when eliminating noise in silent segments,thesis proposes two weak speech enhancement methods from the perspectives of diversity combining and gain control.First,from the perspective of diversity merging,a method of using the CEB＿LSTM network combined with the channel attention mechanism is proposed.After the noise reduction processing of each branch voice,the CNN-Attention network is used to combine the speech signals of each branch.The weak speech signal is enhanced by using the diversity gain of each speech signals.The PESQ score of the combined voice output by the network is about 0.06 higher than that of the single-channel noise-reduced voice,and the STOI score is increased by about1%.Through the simulation data and the actual collected shortwave speech signal verification,The network can better preserve the weak voice signal.secondly,from the perspective of gain control,it is proposed to add a gain control network before the CEB＿LSTM network input,which can dynamically adjust the amplitude of each frame of voice according to the input signal,so that the voice signal and background noise has a greater level of discrimination.Through the actual collected shortwave speech signal verification,this method can solve the problem of eliminating the noise of the long silent segment while retaining the weak speech signal.