Griyo T. Information technology for the search of the given fragments in the archive of audio recordings using kd-trees

Object of study - the processing and retrieval of audio recordings in electronic archives; purpose of study is to improve the recall and precision of the results and the search speed in the archive of audio recordings through the development of the new information technology; used techniques of digital signal processing, cluster analysis, mathematical statistics, theory of algorithms, theory of operation research, computer modeling; theoretical results: the model of audio corpus, containing an array of audio files, parameters, metadata and dynamic kd-tree, is first proposed, which allowes to reduce the duration of the audio fragment and implement the information technology of audio recording search with various duration; the method combined search of audio recording by a given audio fragment is first proposed, based on the executing of approximate kd-tree search for a few recordings among the parameters of the reduced dimensionality at the first stage of search, in order on the second one is to make the final selection of the relevant audio recording, that allowes to reduce the search time compared with an exact kd-tree search; the kd-tree search method is received the further development, which, unlike the existing ones, uses the estimation of the measure proximity based on the weighted number of hits into the list of the nearest centroids, thus enhancing the recall and precision of search results; the clustering method of k-means is received the further development, which differs from the existing ones by the improved selection procedure of the vector as a position of the new centroid by a sequential run of k-means, which provides a solution, close to the global minimum of the clustering error. Practical results - the information technology for search of given audio fragments in the archive of audio recording is developed that contains an algorithm and a program for the implementation of the clustering method based on the sequential run of k-means with an improved choice of the vector as a position of the new centroid; the algorithm and program for search of a given audio fragment based on the mean clustering error; algorithm and program of quick search based on kd-trees. Degree of implementation - the results of research are implemented and used in the research and production organization "Institute of Electronics and Communication of Ukrainian Academy of Sciences" and in the educational process in Vinnytsia National Technical University at the Department of Computer Facilities. Scope (industry) of application - specialized systems for automatic search of multimedia information.