Research Of Speech De-noise Technology Based On Denoising Auto-encoder

Speech signal is the most common communication means for human beings.However,in various scenarios such as speech communication and human-computer interaction,speech quality is easily destroyed by a variety of noises.Therefore,it is of great theoretical and practical value to study effective speech de-noising technology.In recent years,with the improvement of computer hardware's computing capability and the development of neural network technology,the application of deep learning technology in speech de-noising technology has received increasing attention.This paper focuses on the research of single channel speech de-noising technology based on automatic noise classification and autoencoder.Based on the research of speech signal processing and neural network algorithm,this paper proposes a single channel speech denoising algorithm based on stacked autoencoder.In order to ensure the inter-frame continuity of the speech signal,the input speech signal is windowed using overlapped window and processed frame by frame.And subjected to the perfect reconstruction constraint,raising cosine is used to design the window function.In order to make full use of the inter-frame correlation of the signal,the algorithm selects the amplitude spectrum of continuous multi-frame speech as the input feature parameter of the autoencoder,and the optimal number of frames to be cascaded is selected through experiments.The structure of the autoencoder is designed,and some methods such as dropout,L1&L2 regularization and early termination to avoid the model overfitting are introduced.Also an automatic noise recognition algorithm based on convolutional neural network is designed to recognize the noise among 15 kinds of noise in Noise X-92 noise library.In order to balance the performance of speech denoising and model complexity,the 15 kinds of noise is classified into 4 categories according to the spectral features of noise.And the autoencoder model for each category of noise is trained.Finally,the noise recognizer based on CNN and the stacked denoising autoencoder are combined.Using the decision result of the noise classifier the model of the autoencoder is selected automatically.At the end,the speech de-noise algorithm is realized.In this thesis THCHS-30 speech library and Noise X-92 noise library is used to test the noise classifier and speech denoising algorithm.The test results show that the noise classifier has a recognition accuracy rate of 98.9% for the 15 kinds of noise and the proposed algorithm has a good performance of denoising.For nosing speech with SNR of-5d B,the proposed algorithm can improve the average Perceptual Evaluation of Speech Quality(PESQ)by 0.7 and the Short Time Objective Intelligibility(STOI)by 0.37.