Singing Voice Separation by Low-Rank and Sparse Spectrogram Decomposition with Pre-learned Dictionaries

摘要

Unsupervised spectrogram decomposition has shown promising results for singing voice separation in recent years. Its basic idea is to decompose the mixture spectrogram into a sparse spectrogram for the singing voice and a low-rank spectrogram for the background music. This approach, however, has two limitations. First, the unsupervised nature prevents the pre-learning of voice and background music dictionaries from widely available, although not related to the song being separated, isolated singing voice and background music recordings. Second, some components of the singing voice (e.g., fricatives) and the background music (e.g., less repetitive background) may not show the preferred sparse and low-rank properties, respectively. In this paper we propose to decompose the mixture spectrogram into three parts: a sparse spectrogram representing the singing voice, a low-rank spectrogram representing the background music, and a residual spectrogram for the components that are not identified by either the sparse or the low-rank spectrogram. Besides, we learn universal voice and music dictionaries from isolated singing voice and background music training data. Finally, we propose an approach named Low-rank and Sparse representation with Pre-learned Dictionaries under the framework of Alternating Direction Method of Multiplier. Experiments on the MIR-1K dataset and iKala dataset show its better performance.