Research And Implementation Of Single-channel Speech Enhancement Based On Deep Neural Network

Speech enhancement plays an important role in voice signal processing whose aim is to restrain noise and extract speech from the noisy data.There is important significance to discuss and study it in both theory and application.The mono unsupervised algorithms,such as Spectral Subtraction,Wiener Filtering,Nonnegative Matrix Factorization,which based on the hypothesis of the independence of noise and speech are mainly typical speech enhancement algorithms.In addition,it is obvious that this kind of algorithms commonly over-estimate or underestimate non-stationary noise accompanied by unstructured information in low signal to noise ratio.The traditional speech signal processing system operates on speech signal sequence serially,which leads to low velocity and little university.Artificial neural network(ANN)simulates human neurons to handle information by means of distribute parallel processing.At the same time,ANN has strong ability of information extraction,self-learning and fault tolerance.The Shallow neural network is not applicable at solving complicated non-linear problems.In recent years,the deep neural network has achieved great successes in speech recognition and image recognition since it was proposed in 2006.Regressive deep neural network will catch the non-linear relationship between the noise and speech and restrain non-stationary noise.Hence scholars explore the speech enhancement based on regressive deep neural network.In this paper,we conduct systematic study on existing deep neural network model,the main work and achievements are as follows :1)We introduce the framework based on DNN for speech enhancement,which includes feature extraction,the design of network model,the reconstruction of speech.There are two stages for training: pre-training and fine-tuning.Test process involves in the reconstruction of speech and evaluation of speech quality.Also,we analysis the disadvantages of existing speech enhancement algorithms based on DNN.2)Since the existing regression deep neural network model is week in generalization and easy to overfit,we put forward a kind of sparse regression neural network model for training.In both the pre-training stage and fine-tuning stage,we combine dropout method and sparse constraint regularization technique.In this way,it enables a better generalization ability.Also,it can keep the structure consistent between the pre-training stage and fine-tuning stage.In addition,it allows the training model of higher fault tolerance.The experimental results show that the improved training model has better generalization ability than the existing algorithm.3)Speech enhancement algorithms based on existing deep neural network cost large storage,affect the application of the embedded platform.In this paper,the weight quantitative method is applied in the regressive problem.We put forward compression method by weight sharing and weight quantization.The optimization measures guarantee accuracy of neural network and reduce the storage at the same time.Experimental results show that under the condition of small accuracy loss,the technology save storage by 2 ~ 3 times.4)Speech enhancement algorithm based on existing DNN retains much residual static noise.This paper presents a neural network based on post-processing algorithm.The improved algorithm draws support from spectrum subtraction to de-noise.The experimental results show that the method can effectively eliminate the steady noise after speech reconstruction.5)Based on the Matlab simulation platform,we design a software prototype system based on the improved sparse regression DNN.We have tests on testing sets and give performance evaluation of enhanced speech by using the prototype system.The results prove the correctness and effectiveness of the proposed methods in this thesis.