Content based music similarity search

Abstract

The size of personal music libraries of the average computer user is constantly increasing due to the advances in compression techniques and reduction in music file size. Managing such large libraries is always getting more and more time consuming. Living in such a fast paced world, users want to find what they need efficiently. The aim of this project is to design a system which provides a similarity measure function which computes a similarity distance measure between two songs. The interest throughout the scientific community for the development of such a system in a reliable manner is constantly increasing. Music similarity is innate in human beings but very difficult to process using a digital machine. There are two main ways of performing music similarity: (i) Music similarity through content analysis and (ii) Music similarity through the community via relevance feedback. This project aims to measure music similarity based only on the content of the song with no external use of meta-data. This gives the great advantage of computing the similarity measure of two unseen songs. Such a complex task is divided into three main subtasks, (i) feature extraction, (ii) feature profiling and (iii) similarity ranking. Feature extraction is performed using a series of digital signal processing techniques using the highly popular and successful Mel-Frequency Cepstral Coefficients (MFCC). Feature profiling will need to be performed in order to reduce the dimensionality of such a high amount of extracted feature vectors. Profiling of feature vectors is done using a Gaussian Mixture Model (GMM) which aims to estimate the distribution of all the feature vectors using a set of multivariate Gaussian distributions. The initial parameter set-up of the GMM is estimated using the k-means algorithm. Further estimation is performed using the Expectation-Maximization (EM) algorithm. The similarity measure between two tracks is then performed by calculating the probability loglikelihood that a set of generated sample points from the first song profile could have been generated from the second song profile. During the extensive evaluation phase, the system has been found to compute extremely interesting results with certain genres of music (like classical and country music). Certain types of genres (like electronic) are more difficult to rank similarity due to the high inconsistencies in the musical timbre distribution. The main problem encountered with such a system is the high amount of processing required to extract and profile feature vectors. For this reason, the testing corpus was kept relatively low (120 songs from 11 genres). Performing evaluation on a larger test corpus would lead to better results since similarity ranking could be compared with more songs. Although the system produces a lot of interesting results, the system also produces results which are considered not so similar by human beings. Music similarity is a relatively new discipline in computer science and thus more extensive research needs to be performed. An interesting situation in which such a system is highly applicable is playlist generation. By specifying a set of criteria which the playlist should include, the system would automatically parse and match the specified criteria with the numerous amounts of tracks found inside today's user collection of music to produce the required playlist.