Digital Signal Processing Reference
In-Depth Information
Chapter 14
Real-Time Detection of Overlapping Sound
Events with Non-Negative Matrix Factorization
Arnaud Dessein, Arshia Cont and Guillaume Lemaitre
14.1 Introduction
This paper presents non-negative matrix factorization techniques for real-time
detection of overlapping sound events.
1
In general terms,
non-negative matrix factor-
ization
(NMF) is a technique for data analysis, where the observed data are supposed
to be non-negative [
1
-
3
]. The main philosophy of NMF is to build up these observa-
tions in a constructive additive manner. Such assumptions are particularly interesting
when negative values cannot be interpreted (e.g., pixel intensity for images, word
occurrence for texts, magnitude spectrum for sounds).
14.1.1 Motivations
The main goal of this paper is to devise a robust real-time system that processes
rapidly the incoming audio stream and detects the presence of multiple sound events
potentially corrupted by noise. Since several sound events may overlap when con-
sidering realistic situations, we cannot use single-source detection techniques such
as a simple spectral template to audio stream correlation. Instead, we rely on NMF
techniques that intrinsically allow to cope with the simultaneity of the detected sound
events.
The sound events considered in this paper can be produced by various kinds
of sound sources such as a polyphonic instrument (e.g., piano), instruments of a
drum kit (e.g., snare), or environmental sounds (e.g., car horn). These sound sources
1
Additional material including sound files described in the paper are available on a companion
website:
http://imtr.ircam.fr/imtr/Real-Time_Multi-Source_Detection
.
(
B
)
·
A. Cont
·
G. Lemaitre
STMS Lab (IRCAM, CNRS, UPMC, INRIA), 1 place Stravinsky, 75004 Paris, France
e-mail: dessein@ircam.fr
A. Dessein
