Audio source separation and dereverberation are topics in the field of computer audition which have been studied in great depth, though new methods continue to be proposed. In many applications, prior knowledge about the source signal and environment can be applied to produce specialized systems which outperform their generic counterparts in the specific application area.
Since the Fall of 2015, I have been conducting research towards the development of a system that specializes in the source separation and dereverberation of percussive musical audio, such as drum kit performances. The motivation for this project is the relative difficulty and expense of obtaining studio-quality recordings of a drum kit performances when compared to other common instruments. This is due to drum recordings requiring an array of microphones in order to capture different drums in isolation, along with an controlled acoustic environment that is either mostly anechoic or possesses a desirable reverberation quality. However, if percussive sources could be easily separated from a single recording, and ambient reverberation could be effectively removed or replaced, all without introducing audible artifacts or distortion, then more professional results could be obtained with a single microphone and arbitrary acoustic environment than ever before possible.
While this goal is not yet realized, I have made a number of advances towards the ideal of high quality monophonic drum source separation and dereverberation. In the Fall of 2015, for University of Rochester’s ECE 472 – Computer Audition course, I presented my work in semi-supervised source separation for drum kit performances. My system used semi-adaptive NMF to match a mixture of drum sounds to templates obtained from prerecorded “sound checks” of each drum in isolation. The associated report and poster have been found below.
From Spring 2016 to the present, in progress towards my master’s thesis at University of Rochester, I’ve been expanding on this work by developing and evaluating algorithms for joint source separation and dereverberation of drum kit performances. Methods employing NMFD and sparse dictionary learning have been investigated along with the novel approach. This page will be updated with a copy of the thesis upon its submission, which is scheduled for September. A separate research paper relating to this thesis is also in preparation for submission to the 2017 IEEE ICASSP conference.