Gaussian Mixture Model for Singing Voice Separation from Stereophonic Music
Minje Kim, Seungkwon Beack, Keunwoo Choi, Kyeongok Kang · 2011
This paper presents an adaptive prediction method about source-specific ranges of binaural cues, such as inter-channel level difference (ILD) and inter-channel phase difference (IPD), for centrally positioned singing voice separation. To this end, we employ Gaussian mixture model (GMM) to cluster underlying distributions in the feature domain of mixture signal. By regarding responsibilities to those distinct Gaussians as unmixing coefficients of each mixture spectrogram sample, the proposed method can reduce artificial deformations that previous center channel extraction methods usually suffer, caused by their imprecise or rough decision about ranges of central subspaces. Experiments on commercial music show superiority of the proposed method. 1.