Research Of Speech Enhancement Algorithm Based On BP Neural Networks

Speech is corrupted acoustically by ambient noise. It will affect voice quality,which reduce performance of speech processing system. In the noise workenvironment, the purpose of using speech enhancement has been to improve voicequality，and make speech processing system to work more efficiently.For purpose of improving spectral subtraction, this paper studies how to combinetraditional spectral subtraction with BP neural networks. It makes the method adapt todifferent environments. First of all, the principles of some typical classic speechenhancement methods are introduced, what are spectral subtraction, filter method,nonlinear processing and adaptive cancellation method. The emphases are spectralsubtraction and its improved version. And then, this paper analysis principle, structure,training of BP neural network and its application. Based on these, Because spectralsubtraction has residual noise, and BP neural network has premium properties likeself-learning, self-organization, self-adapting and fault-tolerance. This paper proposedthe improvement of spectral subtraction can be identified in the following two aspects.Firstly, voice activity detection based on BP neural network can improve the veracityof u/v decision. Then using all surd speech frames estimate noise spectrum, so as toincrease reliability of noise spectrum. Secondly, this paper acquire factor of spectralsubtraction by nonlinear approximation of BP neural networks. It is differently fromtraditional spectral subtraction, which all speech frames use same factor. In addition,third measure add perceptual weighting filter to spectral subtraction, to make bettercomfort of human auditory.After completing the above steps, computer emulation is made. The simulationresults show the validity of this method.