An End-to-end Bone-conducted Speech Enhancement Method Based On Generative Adversarial Networks

With the rapid development of mobile communications,the problem of voice communication in noisy environments has become an urgent problem that needs to be solved.How to remove the impact of background noise on voice communications has attracted widespread attention.Bone-conducted speech technology provides another anti-noise idea.Bone-conducted speech conduct vibration through the human body,and finally to collect signals through highly sensitive sensors.Because of this special conductive property,bone-conducted speech will not be disturbed by noise in the air,which can eliminate the effect of noise to a certain extent.However,human-conducted and air-conducted speech have different properties,so there is a certain acoustic difference between bone-conducted speech and air-conducted speech.The high-frequency information of bone-conducting speech is severely damaged,resulting in poor hearing of bone-conducted speech.The degree of recognition is not high,which seriously affects the application in noise immunity.In order to solve the problem,we proposes an end-to-end bone conduction speech enhancement method based on generative adversarial network.It uses bone conduction speech sampling points as network input and outputs enhanced speech sampling points.This end-to-end model The enhancement can better utilize the internal information of the speech signal and remove the complex feature extraction and feature synthesis speech process.The generator adopts a convolutional coding and decoding architecture.multiple dilation convolution operations are used to extract the features of the network encoding results at different scales and then fuse them to obtain a stronger expression.In order to enhance the enhancement results,a certain improvement was made to the network loss function,and the learning ability and bone conduction speech enhancement ability were improved through adversarial training.The experimental results show that the method has higher speech perception quality and intelligibility compared with other enhanced algorithms,and also uses the ASR recognition rate as the evaluation index of the voice.The improvement of the recognition rate further confirms the effectiveness of the algorithm.