You are here: Home > For Researchers > Projects > Acoustic event detection - Acoustic event detection: sparse signal representations, embedded optimization algorithms, and case studies

Acoustic event detection - Acoustic event detection: sparse signal representations, embedded optimization algorithms, and case studies

From 01-10-2014 to 30-09-2018

Description

This research project is aimed at the development of a generic framework for acoustic event detection based on sparse audio signal representations, and at the design and evaluation of embedded acoustic event detection algorithms. An acoustic event is defined as a short-time segment of an acoustic signal indicating a change in the physical conditions of the sound source generating the signal. Acoustic event detection has applications in speech/music technology, audio surveillance, ambient assisted living, bio-monitoring, etc.

We will explore a novel approach to acoustic event detection which is based on sparse signal representations. Sparse representations have recently become a powerful tool in audio signal processing and rely on the assumption that a given signal frame can be represented by a small number of basis functions (atoms) from a large and redundant (overcomplete) dictionary. We propose the use of a two-class dictionary consisting of two types of atoms: (1) atoms that can efficiently represent steady-state signal behavior and (2) atoms that can efficiently represent transient behavior. The rationale of our approach is that one can deduce whether or not an event occurs in a given signal frame, by quantifying the relative importance of steady-state versus transient atoms in the sparse approximation of that particular frame.

The proposed approach opens a number of valorization perspectives in different audio applications. These perspectives will be explored in the proposed project through a number of case studies, and several valorization-oriented follow-up research projects will be defined in collaboration with key industrial partners.