Research On Indexing Method Of Audio Sample Retrieval

With the rapid development of information technology, multimedia technology andnetwork technology in the twentieth century, the amount of multimedia data is quicklyincreasing. In order to fully utilize the existing resources of the audio information, audiosample retrieval techniques draw more and more attention. Audio sample retrieval isrealized by handing a query audio sample and searching into the audio database withsimilar content, feedback to the user. In application of audio sample retrieval, speed isthe most important performance. In order to improve the speed of audio sample retrieval,this paper research on indexing method, the details are as follows:(1) The recall rate of Philips method is difficult to meet the requirement of audiosearch application and it is directly related to the extraction process of fingerprinting.The speed of Philips method mostly relates with the size of audio database. When thedatabase is large, the retrieval time will increase proportionally. Much of time is used tocalculate error rate for every candidate match position in the index. We proposemethods of fingerprint extraction process optimization and search optimization toaccelerate index searching speed.(2) Philips method extracts binary fingerprint by quantized energy feature with0or1. Taking into account of the adverse effects of irregularity in time domain, it chooses alarge frame overlap and a small frame shift, which results in a large time consumptionon fingerprint extraction. Spectrum centroid is the weighted mean of energy abovefrequency and can reflect the spectrum character better. We propose NSSCFP Methodusing spectrum centroid, with small frame overlap. The method significantly reducesthe time consumption on fingerprint extraction.(3) Indexing method based on hash table is realized by extracting binaryfingerprint sequence shows good performance of speed. In reality, audio data issusceptible to noise and other interference, which leads to the distortion of fingerprint.Therefore, NN search is the best method to be used. However, the structure of hashtable is not suitable for NN search. We proposed an audio indexing method based onbalanced tree of m branches and NN search is implemented by expanding layer nodes.The index has better performance.