Research On Speech Enhancement Algorithm Based On Deep Neural Networks

In speech signal processing system,speech is usually disturbed by background noise,which seriously damages the speech quality and intelligibility.As a front-end processing module,speech enhancement algorithm has become the focus of research for many scholars.A variety of speech enhancement algorithms have been proposed for noise suppression.They mainly include signal processing based methods,statistical model based methods,and model-based methods.Among these methods,for signal processing based methods,spectral subtraction and Wiener filtering are the two most representative algorithms.This kind of approach can only achieve better speech enhancement performance when the background noise is correctly estimated.However,in the real environments,especially under the condition of low SNR,it's difficult for noise to be accurately estimated due to its randomness and mutation.As a result,it leads to a drastic decline in the speech enhancement performance.At the same time,it's easy to introduce "musical noise".For statistical model based methods,it can also achieve better performance under the condition of low SNR.But considering the complexity of the relationship between noise and speech,some assumptions about the independence of signals and Gaussian assumptions about the distribution of features are needed.However,these assumptions are usually ideal,and algorithm performance deteriorates under unknown mismatched noise conditions.For model-based methods,it shows better results under low SNR and complex background noise.Speech enhancement based on DNN(deep neural network)is a model-based method that has emerged in recent years.Relying on excellent abstraction and modeling capabilities for complex features,DNN has led to extensive research in the field of speech signal processing.The speech enhancement method based on deep neural network has almost no premise assumptions,and can learn well the complex nonlinear mapping functions from noisy speech features to clean speech.In this paper,to solve the problems of feature extraction and objective function optimization in speech enhancement based on deep neural network,from two different perspectives,namely acoustic features and training objectives,three kinds of optimized algorithms were put forward.Firstly,the performance of the proposed algorithm was tested by simulation on TIMIT corpus and Noisex-92 noise database.Then,the speech eavesdropping device was used to collect the noisy speech in the real environment,and the proposed three algorithms were measured.Compared with the traditional speech enhancement algorithms as well as the existing popular deep learning based speech enhancement algorithms,the proposed algorithms achieved better results in enhanced speech intelligibility,perceptual effects and speech quality.