Singing Voice Separation from Monaural Music Based on Kernel Back-Fitting Using Beta-Order Spectral Amplitude Estimation

Separating the leading singing voice from the musical background from a monaural recording is a challenging task that appears naturally in several music processing applications. Recently, kernel additive modeling with generalized spatial Wiener filtering (GW) was presented for music/voice separation. In this paper, an adaptive auditory filtering based on β-order minimum mean-square error spectral amplitude estimation (bSA) is applied to the kernel additive modeling for improving the singing voice separation performance from monaural music signal. The proposed algorithm is composed of five modules: short time Fourier transform, music/voice separation based on bSA, determination of back-fitting, back-fitting, and inverse short time Fourier transform. In the proposed method, the Singular Value Decomposition (SVD)-based factorized spectral amplitude exponent β for each kernel component is adaptively calculated for effective bSAbased auditory filtering performance during kernel backfitting. Using a back-fitting threshold, the kernel backfitting process can automatically be iteratively performed until convergence. Experimental results show that the proposed method achieves better separation performance than GW based on kernel additive modeling.