"Text-Informed Singing Voice Separation and Phoneme Level Lyrics Alignment"

Abstract:

Training data for singing voice separation are scarce and music signals are extremely diverse. Therefore, it remains challenging to achieve high separation quality across various recording and mixing conditions as well as music styles. The question arises to which extent singing voice separation can be improved without access to more audio data. I will present my work on using lyrics transcripts as additional information for deep learning based singing voice separation. It includes a joint approach to phoneme level lyrics alignment and text-informed voice separation.

Bio:

My name is Kilian Schulze-Forster, I am a third year PhD student at Télécom Paris (Institut Polytechnique de Paris) working on informed singing voice separation under the supervision of Roland Badeau and Gaël Richard. Prior to that, I studied sound and vibration.

text_informed_singing_voice_separation_KSF.pdf 1.23 MB, downloaded 3 times Polina Proutskova, 2020-09-22 05:11 PM