ICASSP2023: Hyperbolic Audio Source Separation

Project Description:

This project aims at extending our initial work on the use of U-Net towards neural audio coding tasks (HARP-Net): we propose a multi-band (MB) approach to HARP-Net which offers a convenient band-wise entropy control mechanism. We therefore assume that the network's input (i.e. a full-resolution music signal) is composed a multiple spectral sub-bands, each of which denoting different perceptual relevance to the human ear. We show that having independant entropy control over each of these bands' code can allow us to mindfully attribute the total target bitrate allocation according perceptual motivations.

Video Presentation: