Research On Key Technologies Of Acoustic Event Detection In Audio Monitoring System

Acoustic Event Detection(AED)technology is used to detect and determine the sound types contained in the audio signal such as crying,explosion,glass breaking,door knocking,etc.AED is the key technology in the fields of audio monitoring,smart home,robotics,and industrial flaw detection.This thesis aims to improving the accuracy and robustness of the AED system,and studies and explores the applications of deep learning technology to the AED system,including the application of convolutional recurrent neural network to AED and the application of noise suppression to robust AED.First,for the problem of limited performance of AED system with single feature based neural networks,an AED algorithm based on convolutional recurrent neural networks is proposed.The algorithm is based on two parallel deep learning modules,Convolutional and Recurrent Neural Network and Deep Neural Network,which takes Mel-frequency Cepstral Coefficients features and spectrum features as network inputs,respectively,and then integrates the outputs of the two networks through a fully connected layer.This algorithm not only compensates for the shortcomings of poor noise robustness of the AED system based on the MFCC feature,but also solves the severe training data dependence problem of the AED system based on the spectral feature.Experimental results show that the AED system based on the compound neural network has a better accuracy and noise robustness.Second,to address the problem of poor robustness of AED systems under noise interference scenarios,a noise suppression algorithm based on recurrent neural network is proposed.The algorithm adopts GRU units to construct a recurrent neural network to model and predict the audio signal activity and noise spectrum gain.In addition,in order to improve the accuracy of noise spectrum estimation,a deep clustering network is adopted.Compared with traditional noise suppression algorithms that based on spectral subtraction and Wiener filtering,this algorithm not only has better suppression effect on stationary noise,but also suppresses non-stationary noise where traditional algorithms fail.Experimental results show that the noise suppression algorithm proposed in this thesis improves the performance of the AED system.Finally,this thesis summarizes the key and difficult topics of AED research,and prospects of the research trend.It discusses the classifier construction problem,noise interference problem,weak supervision problem and data annotation problem in the AED system,and introduces the research methods and trends for the above problems,which lays the foundation for further research of AED.