Abstract:
Counterfeiting is the process of imitating the voice of a popular artist, with the
intention of selling or passing the imitations as genuine. This research is aimed
at identifying the imitated songs by generating unique digital signatures for each
singer by using songs sung by the artists.
Songs typically contain vocal signals surrounded with instrumental signals. In
order to generate signatures for the voice of a singer, the vocals have to be
isolated. Voice isolation and Artist classification had been addressed as two
different problems throughout the past decades. Our research combines these two
problems, to build a unique signature model for each singer. This study proposes
a technique to isolate vocal signals from instrumental signals by using REPET
filter, Harmonic - percussive source separation, Butterworth band-pass filter and
silence removal. This technique show better results than the prevailing
techniques for vocal isolation. The signatures are then generated by extracting
features from the isolated vocal signals and represented as GMM Models. The
results of this research are evaluated through quantitative and qualitative
approaches by using a sample of songs sung by Sri Lankan artists. The outcome
of this research has demonstrated the possibility of generating digital signatures
for singers by using their songs. The technique proposed has the ability to
distinguish singers having similar voices accurately. Finally, the technique
proposed in this paper could be used to generate unique signatures for singers
who sung similar songs.
