METHOD OF RAPID SEARCH OF THE AUDIO RECORDING FRAGMENT

Abstract

Taking into account large volumes of audio information, stored in musical compositions, rate and reliability of their search is of great importance. The paper describes the method of rapid search of the audio recording fragment with the improved assessment of the proximity degree between the unknown audio fragment and template, that enables to improve the authenticity of the decision making during the search. For compact description of the signal parameters mel-frequency cepstral coefficients are chosen, on their base the body of music compositions parameters is formed as the set of centroids, obtained as a result of clusterization. The notion of the reduced proper distance as the assessment of the degree of the proximity of the unknown fragment of music composition and previously created templates of the audio recordings is introduced. Application of kd-trees for searching acceleration of the unknown fragment in the body of audio recordings is substantiated, basic stages of the search are presented. Different variants of the proximity degree calculation of the unknown audio fragment with audio recording are considered, namely: assessment of the degree of the proximity by the reduced proper distance, assessment of the degree of proximity by the number of hitting into the list k of the closest centroids, assessment of the degree of proximity by the weighted quantity of getting into the list k of the closest centroids. It is shown, that the execution of non accurate but approximate vectors search on the base of kd-tree enables to achieve considerable time saving, but it leads to the reduction of the validity of the searching results. That is why, to reduce the complexity of the calculations, saving the validity of the results it is suggested to perform combined search for large archives of audio recordings, this type of search combines rapid "inaccurate" search with the application of kd-tree of several nearest audio recordings of the body for the set audio fragment at the first stage of the search, among these audio recordings at the second stage by means of the complete oversampling one the nearest is determined. The suggested method enabled to increase the completeness and relevance of the searching results.