Phoneme Level Lyrics Alignment and Text-Informed Singing Voice Separation

The goal of singing voice separation is to recover the vocals signal from music mixtures. State-of-the-art performance achieved by deep neural networks trained in a supervised fashion. Since training data are scarce and signals extremely diverse, it remains challenging achieve high quality across various recording mixing conditions as well styles. In this paper, we investigate which extent can be improved when lyrics transcripts used additional information. To end, propose joint approach phoneme level alignment text-informed separation. It based on DTW-attention, new monotonic attention mechanism including differentiable approximation dynamic time warping. Experimental results show that method align phonemes with mixed precision given accurate transcripts. also achieves competitive word test sets using less than state-of-the-art methods. Sequential informed lead according objective measures. Text information helps preserving spectral properties separated signals.