Research On Audio Classification And Recognition Based On Sparse Representation And Topic Model

Audio information is an important means for people's perception of the external environment.It can assist visual information in situations such as eye-shielding,poor lighting conditions,and privacy occasions.It also has the role that visual information can not be replaced.With the rapid development of multimedia information and the rapid increase in the amount of audio information,the demand for audio information management and applications is increasing.The research of audio information has received more and more extensive attention.Audio information has a wide range of application prospects,such as audio emotion perception,smart home engineering,and scene recognition based on audio information.Audio event classification and audio scene recognition are two important research directions in the field of audio classification.In recent years,researchers have received extensive attention.This thesis researches on audio event classification and audio scene recognition based on sparse representation technology and topic model technology.The main research work includes:(1)This paper proposes an audio event classification method based on sparse base sparse representation.This method uses the K-SVD algorithm to train and create an audio dictionary for each type of audio event.After obtaining the basis functions for each type of audio event,a large audio dictionary is obtained by stacking various types of basis functions,and finally based on the new The created large audio dictionary extracts sparse representation features of the audio signal.In the classification stage,this paper proposes a classification strategy based on calculating the weight values of samples on various types of audio events,and then discriminating based on the weight values.When the speech-music two-class classification experiment was conducted,the classification method proposed in this paper has a classification accuracy rate as high as 100%.When the speaker recognition experiment was performed on the TIMIT database,the classification accuracy rate was as high as 95%,which was 13% higher than the maximum pooling sparse method proposed by Syed Zubair of the University of Surrey in the United Kingdom.(2)This article proposes an audio scene recognition method based on audio events and topic models.Different from the traditional document-word co-occurrence matrix for thematic analysis,the algorithm proposed in this paper performs thematic analysis by creating audio document-audio event co-occurrence matrix.The innovation of the algorithm is embodied in : 1)Compared with the traditional thematic analysis method based on document-word co-occurrence matrix,the topic analysis based on audio document-audio event co-occurrence matrix can better extract the topic distribution of audio files and better express audio.Document,and then get a better recognition effect;2)proposed a simple statistical method of audio files-audio event co-occurrence matrix;3)propose a method to weight the event distribution of audio documents.The method,this weighting method can highlight important audio events that reflect the unique theme of the audio document,and can suppress audio events common to many topics.Experiments on AASP database and DEMAND database show,in the recognition performance,the audio scene recognition method based on audio document-audio event co-occurrence matrix proposed in this paper is superior to the traditional audio scene recognition method based on audio document-audio word co-occurrence matrix.