Research On Single Channel Speech Enhancement Based On Combined Convolution

The purpose of speech enhancement is to separate clean speech from speech signal polluted by additive noise,and then improve the clarity and intelligibility of speech.In the past,the traditional speech enhancement methods based on signal processing have good performance on noise reduction for stationary noise.But when the surrounding noise environment becomes very bad or no longer stable,the filtering effect of noise decreases sharply and results in serious speech distortion.In recent years,speech enhancement algorithm based on deep learning has become the mainstream method in the field of speech enhancement because of its remarkable enhancement effect under the background of low SNR and complex noise.Among them,convolutional neural network can obtain excellent noise reduction effect under the condition of using fewer model parameters,which is favored by scholars.On this basis,this paper carries out relevant research,the main content and innovation points are shown as follows:(1)The extensive use of dilated convolution in convolutional neural network has insufficient ability to extract information of adjacent points,which affects feature extraction of network.Further it limits the improvement of speech enhancement performance of the model,and results in bad effect on the reconstruction of original speech.In order to solve the problems above,a combined convolutional neural network for single-channel speech enhancement named CB-ConvNet,was proposed in this paper.By using two parallel convolutional blocks to extract and merge the long distance information and short distance information of feature maps respectively,the speech enhancement performance of the model was improved.(2)In this paper,it is found in the experiment that the combined convolution block in CB-ConvNet has a certain feature loss in the process of extracting short distance information.To solve this problem,this paper also proposes an adaptive squeeze convolution neural network named ASConvNet based on sub-pixel convolution.It can extract multi-scale short range information and reduce feature loss.Compared with CB-ConvNet,AS-ConvNet further improved the noise reduction performance of the model by adding a few parameters.