Study On Speech Enhancement Algorithm Based Deep Neural Network

A great many types of noise are ubiquitous in the actual environments. Speech signals for human communication are seriously polluted by noise. There are a lot of speech enhancement algorithms at present, but in the low Signal-to-Noise Rate(SNR) complex noise environments, the existed approaches could not fulfill the requirements of real-world applications.In view of the limitation of speech enhancement methods in low SNR complex noise environments, we focus on speech endpoint detection algorithm based on BP neural network and multiple features, Deep Neural Network(DNN) based noise classification method and Improved Least Mean Square Adaptive Filtering(ILMSAF) based speech enhancement approach with DNN and noise classification. The main research contributions consist of three parts:1. In order to improve the performance of speech endpoint detection algorithm in low SNR complex noise environments, a speech endpoint detection method based BP neural network and multiple features is proposed. Firstly, taking into account the time domain feature and frequency domain feature of speech signal, maximum of short-time autocorrelation function and spectrum variance are extracted. Secondly, the two features as 2-dimensional vectors are inputted to BP neural network to be trained and modeled. Then the parameters of BP neural network are optimized by Genetic Algorithm. Finally, according to an adaptive threshold, the type of the current frame signal is determined by the trained BP neural network. The experiments show that compared with single feature a nd linear model, the proposed algorithm has better adaptability and robustness in low SNR complex noise environments. Meanwhile, the accuracy of speech endpoint detection is also further improved.2. Focusing on that the different kinds of noise signals ha ve different effects on speech signal, in order to make the speech enhancement algorithm suitable for various kinds of noise environments, a noise classification method based on DNN is presented. Firstly, the Mel Frequency Cepstrum Coefficient(MFCC) and first-order MFCC(?MFCC) of non-speech segments in(1) are extracted respectively. Secondly, the MFCC and ?MFCC as 24-dimensional vectors are inputted into DNN to be trained and modeled. Finally, the type of the current frame noise signal is determined by t he trained DNN. The experimental results indicate that compared with Gaussian Mixture Model(GMM) based noise classification, the classification accuracy of the presented method is improved.3. Focusing on some shortcomings of the existing speech enhanceme nt algorithms, for example, non- ideal performance in low SNR environments, poor adaptability in various types of noise environments and the difficulty to process nonstationary noise, an ILMSAF speech enhancement model is developed. Through the introduction of an adaptive coefficient into the traditional LMSAF speech enhancement, the filter parameters are updated, which make filter more effectively remove the noise. Based on that, an ILMSAF speech enhancement approach with DNN and noise classification is proposed. First, the adaptive coefficient of filter's parameters is estimated by Deep Belief Network(DBN). Then, the enhanced speech is obtained by ILMSAF. In addition, according to the result of noise classification in(2), the corresponding ILMSAF model is trained, which makes the speech enhancement algorithm suitable for various kinds of noise environments. The performance test results under ITU-TG. 160 show that, the proposed algorithm tends to achieve significant improvements in terms of various speech subjective and objective quality measures than the wiener filtering based speech enhancement approach with Weighted Denoising Auto-encoder and noisy classification.