Design & Implementation Of Medolic-Based Music Retrieval System

This thesis studies the performance issue often encountered in the context of similarity searching within content-based music retrieval system having large-scale collection of pieces of music. In an attempt to solve this issues, a technology called Locality Sensitive Hashing has been examined and used to index music melody collections to speed up the process of query, to this end, a full featured prototype of Content-Based Music Retrieval System has been developed to facilitate the evaluation of its effectiveness.Generally, Music Retrieval System with its query criteria based on melodic characteristics allows user to submit a piece of digital audio content as query condition, upon received, a melodic contour has been extracted from the query condition, then the extracted melody contour will be compared with each one in the music collections to measure their similarity, finally top-n most similar list of pieces of music will be returned to the user as the result of search. While being simple to implement, this approach has inherent problem of being unable to scale to large collection of pieces of music due to the fact that it has to do the time consuming similarity computing between each melody contour with the counterpart from the query, as a result, the query response will degrade linearly as collections of music pieces growing, in order to retain relatively stable performance of query processing within system having large collection of pieces of music, one feasible way is to represent melody contour as a point in metric space, then index the collections of melody counters using the underlying metric distance. with the index being constructed beforehand, upon a query request has been received, it can be used to filter out those melody contours which are not possible to be similar to the query so after this stage, only small collections of candidate melody contours survive to the next stage of exact similarity computing. The similarity computing used in next stage is dynamic time warping, a bit of time-consuming though, deliver a good ranking quality. Since the number of candidate melody contours is relatively small and does not depend on the size of collection of pieces of music, a good response of query processing is promising.In particular, a Locality Sensitive Hashing scheme with its distance measure as Euclidian distance has been used to construct the index of collection of melody contours, then the query processing has been divided into two stage, at the first stage, Locality Sensitive Hashing values have been computed against the query contour and then are used to index the candidate melody contours, at the second stage, each melody contour within the candidate melody contours obtained in the first stage will be compared with query melody contour to work out the final rank list of similar melody, once it has been finished, a top-n most similar list of pieces of music will be returned to used as the result of query. The result of the prototype shows that it can deliver promising query performance.