Ofer Hadar

This paper examines the use of the generalized likelihood ratio test (GLRT) for the purposes of audio extraction and manipulation. The GLRT, which was designed originally for distinguishing between harmonic and non-harmonic audio frames, is extended for two music-oriented purposes. The first is to decompose a multiple-source mono recording into separate sources by which the decomposed files may be used to create new interpretations of the original recording. The second is for the purpose of lyrics alignment. The test shows a clear distinction of the singing voice within an orchestrated recording. Furthermore, words and syllables are indicated and can be used to align lyrics to the music automatically.