Mixing or Extracting? Further Exploring Necessity of Music Separation for Singer Identification

One song has two major acoustic components that are singing vocals and background accompaniment. Although identifying singers is similar to speaker identification, it is challenging due to the influence of background accompaniment on the singer-specific information in singing vocals. In past work on singer identification, studies on smaller datasets have considered the introduction of audio-source separation to remove the accompaniment to be beneficial for singer identification. In our work, this was not found to be absolutely valid for identification accuracy on a larger dataset with a wider variety of acoustic environments. Moreover, to further illustrate the necessity of removing accompaniment in the singer identification problem, we collected three characteristic datasets focusing on backing tracks for publicly released songs, cover songs, and multiple songs per singer. And general and specific system performance example results are given to reveal the effectiveness and reliability of removing the accompanying sound.