Research And Implementation Of Lightweight Speech Enhancement Algorithm For Air Control

Speech is the most common form of communication.In order to make machines replace humans,developers have come up with speech recognition technology,which enables machines to replace human brains to convert speech into text.However,the accuracy of speech recognition algorithm in a noisy environment will drop significantly,so it is often necessary to use speech enhancement technology as its front-end for processing.In this thesis,a lightweight speech enhancement scheme for complex noisy environments is explored.An algorithm that can adaptively learn the noise characteristics under different noisy environments and perform noise reduction in a targeted manner is studied.Since the proposed algorithms are all based on convolutional neural networks,different cutting-edge lightweight convolutional layers are introduced into the model to replace the original ordinary convolutional network layers to achieve model compression,which is convenient for better application in embedded and other small in the device.In addition,the proposed network model is also applied to a real project.Overall,the contributions of this thesis are as follows:(1)Come up with noise-adaptive speech enhancement algorithms.For the great generalization ability of the model,so that the model can be trained in a limited noise data set,and can still achieve a good noise reduction effect when tested in an infinite noise data set,consider introducing the squeeze-and-excitation module into the encoderdecoder models based on convolutional neural networks.In this way,the model can adaptively filter noise features and retain speech features,thereby effectively dealing with different noises and achieving good performance under complex noise conditions.(2)Transform the compression excitation module,so that the model can be adapted from different angles.In aim of exploiting the extra features except for channel information,the transformation operation is performed on the compression excitation module.The original compression excitation module is to assign a weight to each channel.In this thesis,a new time-frequency compression excitation module is proposed to assign a weight to each time-frequency feature vector.And the compression excitation modules of these two different dimensions are combined in various ways to maximize the use of the feature information in the audio signal.(3)In order to effectively apply the model to small embedded devices such as telephone devices,this work uses a cutting-edge lightweight convolutional neural network to replace the original ordinary convolutional neural network.On the premise that the noise reduction performance is only slightly reduced,it is effective.Reduce the number of parameters of the model,thereby improving the applicability of the model.(4)In aim of exploiting the models in the actual scene,and clearly show the effect of the model,this work builds a system for display,and the proposed algorithms were integrated into it for real-time display.