eduzhai > Applied Sciences > Engineering >

Adversarially Trained Multi-Singer Sequence-To-Sequence Singing Synthesizer

  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: This paper presents a high quality singing synthesizer that is able to modela voice with limited available recordings. Based on the sequence-to-sequencesinging model, we design a multi-singer framework to leverage all the existingsinging data of different singers. To attenuate the issue of musical scoreunbalance among singers, we incorporate an adversarial task of singerclassification to make encoder output less singer dependent. Furthermore, weapply multiple random window discriminators (MRWDs) on the generated acousticfeatures to make the network be a GAN. Both objective and subjectiveevaluations indicate that the proposed synthesizer can generate higher qualitysinging voice than baseline (4.12 vs 3.53 in MOS). Especially, the articulationof high-pitched vowels is significantly enhanced.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×