Audio Classification And Segmentation Method Based On Beats And Key Background Models

Audio data classification and segmentation are the core of numerous applications and the core part that effect the performance of the subsequent transaction processing,so the research of this technology has been paid extensive attention to.In the audio sample retrieval system,it is necessary to provide users with a convenient and efficient sample extraction tool,which intercepts the audio clips from the multimedia files as the sample data of the query.Examples of interception are usually contain the contents of important information which are interested to the users and are generally voice information.Although the audio information can not quickly browse,but if before the user to capture audio sample,structured information first acquiring audio files using the audio content segmentation and classification technology,and then the visualization technology is presented to the user,to facilitate users to quickly browse the contents of the file,convenient and efficient selection of audio c lips,complete sample production tasks.In this paper,the audio sample extracts as the application background,the research and implementation of audio data classification and segmentation algorithm based on multi-level classification,and the algorithms used for sample extraction to improve the work efficiency of operating personnel,the specific work is as follows:(1)The coarse classification of audio is realized by random forest algorithm,and audio data is divided into five categories: music,speech,speech with music,speech with noise and background sound.And audio data is divided into fragments of three types: music,speech and background sound.(2)For the music part after roughly classified,because singing is more important than pure music,this paper puts forward a classification method based on music beats to discriminate pure music between singing.The experiment shows that the proposed method for different type of data were obtained improved performance.(3)For the part of speech after roughly classified,if the speech is too long,it is not convenient to quickly select the extraction location of the sample fragments.A speech segmentation method based on gaussian statistical histograms feature and single gaussian-key background model is proposed in this paper.Train model and select the best models,using statistical methods for mapping Mel cepstrum coefficients feature into histogram feature.And then we use the clustering algorithm and cosine distance to cluster data and choose the number of clustering function to get the optimal clustering number then we get the clusters.Using the clusters to divide the speech into small independent audio clips.Experimental results show that the proposed speech segmentation method has ideal accuracy and low computational time.