Research On Sound Scene Classification In Intelligent Digital Hearing Aid

With the continuous development of society,the aging problem is more and more prominent,which prompted people to pay attention to the problem of deafnesss,and the development of hearing aids begin to receive people's attention.As the core technology of intelligent digital hearing aids,the sound scene classification algorithm is at the front of signal processing,it can identify the current auditory scene of the hearing aid's user,adaptively call the corresponding processing program,and realize the personalized processing for different scene sound signals.In essence,sound scene classification is an environmental sound recognition problem,which mainly includes two aspects:feature extraction and classification.Feature extraction is the dimension reducing of the sound signal,and the data which can represent the original signal is extracted.Classification means that the acoustic feature is encoded by a certain method,and compared with the template database to determine the category of the sound signal.This paper mainly focuses on selective attention model,traditional HMM model and deep learning model,the core issue is the study of sound scene classification.In the aspect of feature extraction,the paper makes a significant analysis to the spectrum of the sound signal,extracts the significant feature,and then mixes it with the traditional MFCC feature to form the mixed feature.In the aspect of classifiers,the paper uses the traditional HMM model and the deep learning model to classify.The paper mainly completed the following works:1.The research background and research status of sound scene classification technology are expounded,the advantages and disadvantages of the existing technology are analyzed,and the problems that need to be studied and solved are explained.2.The paper introduces the basis theory of sound scene classification,introduces several modules of sound scene classification,including preprocessing module,feature extraction module,classifier training module and test module,and introduces the characteristics of sound scene classification feature extraction and classifier design related information.3.In the paper,the selective attention model is studied,including the theory of selective attention model and two kinds of commonly used models--the Itti model and the GBVS model,the visual theory is applied to acoustics,the saliency analysis of the spectrum is finished,and the saliency map parameters of the sound signal are extracted.4.The feature vector for classification is obtained by extracting the saliency map parameters,and then the MFCC features of the sound signal are extracted,and the two are mixed to form the mixed feature.Then the traditional HMM classifier respectively using separate saliency map characteristics,separate MFCC features and mixed finish for the classification of the sound scenes,and comparing the classification results.5.The paper introduces the development,the main models and the applications of deep learning,and introduces the commonly used methods of deep learning,including automatic encoder,sparse automatic encoder and restricted Boltzmann machine,and the Gibbs sampling and the Contrastive Divergence algorithm are also introduced.In the paper,deep learning model is applied to sound scene classification.Two kinds of commonly used deep learning models--sparse automatic encoder and deep belief network are used to compose the mixed model.The model consists of three parts:two layers of sparse automatic encoder are used in front of the system,and three layers of deep belief network are used in the middle,and Softmax regression is used as a classifier.Then,using the saliency map faetures,MFCC features and mixed features to experiment,and comparing the respective classification results.