INVESTIGATION OF IDENTIFYING COVER SONGS FOR SINHALA SONGS UNDER MUSIC INFORMATION RETRIEVAL

Abstract

Cover songs are alternative versions of previously recorded songs. Simply versions. Nowadays versions have become big ethical issue around copyright enforcement of artists in Sri Lankan music context. In this situation, we can identify the necessity of an appropriate solution for this problem. But it is a challenging task in local music context. Our music context has influenced by other music cultures and there are more types of cover songs and various features among cover songs. Then we proposed a novel and hybrid approach which followed state of art approach in cover song identification. The approach consists two sections. Song registration and version identification. We can mostly find mp3 format of Sinhala songs and then we prepared them for registration after removing repeated melody parts and checking the channel. After this pre- processing, then we extracted features of each mp3 to Chroma matrix while storing Chroma feature matrix in a directory for cross-correlation. Then we converted each beat synchronous Chroma matrix to a double vector using novel key invariance method. After that, vectors were stored in the database. In next section, extracting features and key invariance are same as previous section for a given input mp3. Similarity computation consists of two filers. Filter 1 extracts songs those have a cosine similarity value greater than the threshold value. Then extracted songs are sent to filter 2 and there, identified higher cross-correlation coeficients greater than to the threshold for song execution. According to this methodology we implemented the system and evaluated under performance and accuracy. One song is taken 3.93s average time for song registration. Also average execution time of a song is 6.53s. System has been evaluated using 10 test cases and we have identified that tempo, live performance and structured varied songs cause to reduce accuracy. But all test cases showed accuracy greater than 75%. When we take overall accuracy, it is 88% and it is a good achievement while comparing with existing approaches. According to the evaluation, we could have provided proper solution to identify cover songs of Sinhala songs. Here we considered only melody related parameters on versions which have clear and related melodies for correct identifications. Some of suggestions and weak points will be considered in future works.