4 Maig - Presentació Marc Ferras
As some of you already know, Marc Ferràs is doing a postdoctoral stay with us till next July.
Taking advantage of this opportunity, we are glad to invite you to his oral presentation on 'Hiercarchical Adaptation for Speaker Recognition'.
When: Wednesday 4th May at 11:15
Where: 'Sala de Juntas' (D4, ground flour)
Short Bio
Marc Ferras has recently finished a postdoctoral visit at the Furui Laboratory (Tokyo Institute of Technology, Japan) where he has researched the discrimination and separation of speaker and session effects in speaker recognition tasks. He got his PhD. degree at the LIMSI-CNRS laboratory (Orsay, France) in 2009, also in the field of speaker recognition. In the past, he has worked on a variety of topics from mobile robotics to speech enhancement and recognition, at the International Computer Science Institute (ICSI, Berkeley, CA) and several laboratories of the Technical University of Catalonia (UPC, Barcelona, Spain). His current interests are speaker and speech recognition and machine learning.
Abstract
In this talk I will discuss several variants of Maximum A Posteriori (MAP) adaptation of Gaussian mixture models (GMM) and how speaker and session effects can be handled in each of them for speaker recognition tasks. I will revisit relevance MAP adaptation as well as less standard techniques such as Structural MAP (SMAP) adaptation and the successful Joint Factor Analysis (JFA), focused on speaker and session effects separation. I will also introduce a hierarchical version of JFA, i.e. Structural JFA (SJFA), as a hybridization of the SMAP and JFA frameworks for improved performance in scarce data scenarios that still require separate adaptation to speaker and session variation. I give results for all of these methods on a speaker recognition system using GMM-supervectors and Support Vector Machines evaluated on the 2006 NIST Speaker Recognition Evaluation data.