Project Description:

My ongoing internship with Mitsubishi Electric Research Laboratories (MERL) led to the public release of a dataset addressing the novel task of music, speech, and sound-effects source separation. The "Divide and Remaster" (DnR) dataset is proposed as part of the Cocktail Fork Problem, in which we present and investigate this task. Link to the project page, examples, code repo, and ICASSP paper are provided above. In the demo below, we show you how our proposed model, trained on the DnR dataset, performs on a real-world soundtrack. We dynamically interpolate between the three sources of interest using a custom mixing interface developed for this project.

Video Presentation:

Model in Action: