Fusion of Classical Digital Signal Processing and Deep Learning methods (FTCAPPS)

Abstract

The use of deep learning approaches in Signal Processing is finally showing a trend towards a rational use. After an ef- fervescent period where research activity seemed to focus on seeking old problems to apply solutions entirely based on neu- ral networks, we have reached a more mature stage where in- tegrative approaches are on the rise. These approaches gather the best from each paradigm: on the one hand, the knowledge and elegance of classical signal processing and, on the other, the great ability to model and learn from data which is inherent to deep learning methods. In this project we aim towards a new signal processing paradigm where classical and deep learning techniques not only collaborate, but fuse themselves. In partic- ular, we focus on two objectives: 1) the development of deep learning architectures based on or inspired by signal processing schemes, and 2) the improvement of current deep learning train- ing methods by means of classical techniques and algorithms, particularly, by exploiting the knowledge legacy they treasure. These innovations will be applied to two socially and scientifi- cally relevant topics in which our research group has been work- ing for years. The first one is the enhancement of speech signal acquired under acoustic adverse conditions (e.g., noise, rever- beration, other speakers, ...). The second one is the develop- ment of anti-fraud measures for biometric voice authentication, in which banking corporations and other large companies are strongly interested.