LSP external seminar series

A brief introduction to deep learning and its application to multichannel speech enhancement

Informations pratiques
07 mars 2018
11h-12h
Lieu

Salle Paul Langevin, 2ème étage, 29 rue d'Ulm, 75005 Paris

LSP

Over the past decade deep learning has become the state-of-the-art in many applications including several tasks of speech and audio processing. It has recently been applied to multichannel speech enhancement, outperforming most of the classical approaches. In this presentation, I will present a brief history of neural and a short overview of some deep learning architectures that are currently used. I will then describe the problem of multichannel speech enhancement and a solution to this problem: the multichannel Wiener filters. Finally, I will present recent works that capitalize on both multichannel filters resulting from decades of work in signal processing and on the modeling power of deep learning to design deep learning based multichannel speech enhancement algorithms that are now the state-of-the-art in the domain.