Speech Enhancement Based On Linear Prediction And Generative Adversarial Network For Bone-conducted Speech

Speech is a very primitive but efficient way of communication.However,speech communication is always disturbed noises in realistic environment.Noise will hinder the listener from receiving voice information efficiently and accurately.Researchers have proposed many methods to solve the noise problem in speech,but few methods can achieve satisfactory results in extremely noisy environments.Recently,the demand for voice communication in highly noisy environment has been increasing.Speech transmitted on bones is called bone-conducted speech.When voice waveforms are transmitted through the vocal tract and skull,they do not directly meet with noise.That is why the bone-conducted speech is used to accomplish speech communications in a very noisy environment.It can satisfy the needs in special situations,but its quality and intelligibility are lower than air-conducted one as it faces severe degradation due to the attenuation of human body channel.Therefore,some methods should be taken to get a better application of bone-conducted speech.This paper proposes a method of bone-conducted speech enhancement based on generative adversarial network,and implements it with G.729 codec.The main contributions of this paper are as follows:(1)The method of bone conduction speech enhancement is combined with G.729 codec in communication,the combination will make the application of boneconducted speech equipment more practical.(2)The previous method of bone conduction speech enhancement based on linear prediction model is optimized,new features are extracted to train the model and converse.Comparative experiments show that our improvement can achieve better performance of bone-conducted speech enhancement.(3)A network model combining long short-term memory network with generative adversarial network is proposed for bone-conducted speech enhancement task.The experimental results show that the network has better effect than others.