< Back to previous page

Project

Learning signal representations of acoustic manifolds

Efficient analysis and processing of temporally and spatially sampled sound fields is paramount for many applications, including acoustic scene analysis, room geometry inference, and acoustic signal enhancement. Traditional approaches often rely on simplified physical and statistical models of the data and the data generation processes. With the advances in machine learning, prior models are being replaced by models learned from training data, often using deep neural networks. This approach has proven very successful for speech recognition and for single-channel signal enhancement. However, we believe that in the case of spatio-temporal acoustic signals, the underlying physical properties of the acoustic channels convey relevant geometric information which can be used to enhance the discriminative information contained in the signals, more efficiently than in purely data-driven approaches. Hence, our main objective is to develop a generic approach to learning data-driven representations of multi-channel audio signals that exploits the geometric and physical properties of the sound field and the acoustic channels. We believe that by imposing meaningful geometry-related constraints in the representation-learning optimization problems, we can derive signal representations that would pave the way for more efficient semi-supervised approaches to acoustic signal analysis and enhancement, even with a limited amount of training data.
 

Date:1 Oct 2018 →  31 Aug 2019
Keywords:acoustic manifolds, learning signal representations
Disciplines:Sensors, biosensors and smart sensors, Other electrical and electronic engineering