Thursday, February 12, 2015, 14:00 pm to 15:00 pm, room F107, INRIA Montbonnot
Seminar by Romain Sérizel, Télécom ParisTech
A brief introduction to deep neural networks and their application to automatic speech recognition
Abstract. During the past decade, with advances in terms of training algorithms and computing power the deep neural networks (DNN) have become the state-of-the-art approach in many applications related to signal processing. The DNN owe their increasing popularity to their good generalization capabilities, the discriminative training, the use of extended context without loss of temporal resolution and the fact that their deep architecture allows them to extract high level features invariant to small variations. During this seminar, I will briefly introduce the DNN and theprogression leading from the perceptron model in the late 50’s to the now commonly used, deep architectures and (pre)training procedures relying for example on back-propagation and restricted Boltzmann machines. I will then describe an application of DNNs to automatic speech recognition where they recently allowed the emergence of reliable mass market products. Finally, I will present experimental results on a challenging scenario that is automatic recognition of children speech.
