A Singing Voice/Music Separation Method Based on Non-negative Tensor Factorization and Repeat Pattern Extraction

摘要

In this paper, a novel singing voice/music separation method is proposed based on the non-negative tensor factorization (NTF) and repeat pattern extraction technique (REPET) to separate the mixture into an audio signal and a background music. Our system consists of three stages. Firstly, we use the NTF to decompose the mixture into different components, and similarity detection is applied to distinguish the components from each other, in order to classify the components into two classes as the voice including voice/periodic music and the block music/voice; next we utilize the REPET to extract the background music one step further for the two classes, and the fined background music is estimated by adding the two backgrounds together, the left is added together as the singing voice; finally the music spectrum and the voice spectrum are filtered by harmonic filter and percussive filter respectively. To improve the performance further, wiener filter is used to separate the voice and music. Our method can improve the separation performance compared with the other state-of-the-art methods on the MIR-1K dataset.