Research On Speech Enhancement Algorithm Based On Convolutional Neural Network

Today,artificial intelligence is more and more mature and voice technology has become a key interface for human-computer interaction.However,various background noise in real life seriously interferes the voice interaction.In order to improve the voice interaction effect,voice enhancement is needed to filter out the noise signals in the voice interaction so as to improve the clarity,comprehension and fluency of the voice communication.In traditional speech enhancement algorithms,spectral subtraction has the defect of "music noise".Other algorithms also have some disadvantages such as difficulty in dealing with non-stationary noise and so on.In view of this,this paper make a reacher on speech enhancement algorithm based on smooth fast recursive least squares(SFTRLS)and convolutional neural network(CNN).The main contributions of this dissertation are as follows:(1)A noise identification algorithm based on CNN is proposed.The algorithm has different influences on different kinds of noise and identifies the type parameters of environmental noise,which makes the enhanced model suitable for different noise environments and improves the adaptive ability of the algorithm.Experimental results show that the noise identification rate of CNN is up to 99.97%.Compared with the noise identification algorithm based on KNN model and support vector machine,the noise classification algorithm proposed in this paper has better accuracy.(2)Proposes a speech enhancement model combining CNN and SFTRLS ——SFTRLS-CNN.The model takes the output of noise recognition as input.By using the characteristic of SFTRLS algorithm that the convergence performance and the non-stationary noise suppression are better than the traditional algorithm but strongly dependen on forgetting factor,the model learns the best forgetting factor of SFTRLS for each noisy environment through offline training.In practical application,the algorithm matches the best forgetting factor coefficients and then enhances the noisy speech through SFTRLS.Experiments show that the algorithm can achieve an accuracy of 99.40%.Compared with the common SFTRLS,the algorithm can achieve better PESQ(Perceptual evaluation of speech quality)and the degree of distortion is less.