Bilgisayarlar ve Ses Tanımada Gelinen Son Nokta

TalkPrinting: Using Advanced and Stylistic Features for Speaker ID

Abstract:

Speaker recognition has seen significant new developments in recent years, as reflected by the performance of the systems at the NIST Speaker Recognition Evaluations (SRE) of the past few years. In this talk, I will give a brief overview of the current state-of-the-art then describe our approach, /TalkPrinting/, that involves a rich stylistic representation in addition to several novel acoustic ones. SRI's /TalkPrinting/ systems has been the top performing system in 2004 and 2005 at NIST SREs. It involves seven subsystems that account for speaker variation at different scales. I will introduce each system and also talk about system combination and describe statistical techniques for measuring and optimizing the final performance. Time permitting, I will also briefly show SRI-Speaker Recognition Dashboard, a SRE system browser that highlights how combination of stylistic and acoustic features helps by displaying intermediate features and models as well as the final scoring in a verification system.

Short Bio:

Kemal Sönmez received his Ph.D.in Electrical Engineering from the University of Maryland College Park in 1998. He joined the Speech Technology and Research (STAR) Laboratory of SRI International (former Stanford Research Institute) in 1996 and has been a senior research engineer at SRI since 2001. He has been PI or key technical member on several speech and speaker recognition, and biocomputation programs, including DARPA programs ROAR, EARS, GALE, BioCOMP, and NSF program KDD, as well as, more recently, several NIH funded projects in computational biology. He was a member of the Institute for Mathematics and its Applications (IMA) Fall 2000 Workshop on Mathematical Foundations of Speech Processing and Recognition at the University of Minnesota, a visiting research scholar at the Division of Applied Mathematics and Center for Computational Molecular Biology at Brown University in Providence during the Spring 2004 semester, and a senior member of the landmark based speech recognition team at the 2004 Center for Language and Speech Processing summer workshop at Johns Hopkins University in Baltimore. In addition to topics in speech and language processing, his research interests include statistical modeling of evolution in genomic sequences, signal transduction models, discovery of orphan GPCR ligands, pathway models, and building integrated models for systems biology. He was the proposal manager and a system architect for SRI-UC Berkeley consortium for BioSPICE, an open source systems biology platform supported by DARPA and managed by SRI.