Application of SHAZAM-Based Audio Fingerprinting for Multilingual Indian Song Retrieval

Abstract

Extracting film songs from a multilingual database based on a query clip is a challenging task. The challenge stems from the subtle variations in pitch and rhythm, which accompany the change in the singer's voice, style, and orchestration, change in language and even a change in gender. The fingerprinting algorithm must be designed to capture the base tune in the composition and not the adaptations (or variations which include lyrical modifications and changes in the singer's voice). The SHAZAM system was developed for capturing cover audio pieces from millions of Western songs stored in the database, with the objective of tapping into the melodic construct of the song (devoid of other forms of embellishments). When applied to the Indian database the system was found less effective, due to subtle changes in both rhythm and melody mainly due to the semiclassical nature of Indian film songs. The retrieval accuracy was found to be 85 %. Potential reasons for the failure of this SHAZAM system have been discussed with examples.