学  术  报  告

报告题目：Recent Advances in Discriminative Training and Adaptation Techniques for Large Vocabulary Continuous Speech Recognition

报告人：  征荆博士

 Speech Technology and Research Laboratory, SRI International

报告时间：2007年12月6日 上午10：00

报告地点：FIT楼1-312

主办单位：电子工程系

联系人：  丁晓青教授

简介：

During the past few years, discriminative training and adaptation techniques, such as minimum phone error (MPE) and maximum mutual information estimation (MMIE) training, as well as their maximum a posteriori (MAP) variants, have shown great effectiveness in improving recognition accuracy in various large vocabulary continuous speech recognition tasks, over the traditional maximum likelihood estimation (MLE) based training and adaptation. In this talk, I will present our works on advancing the standard discriminative training techniques in the aspects of improving objective function, speeding up training procedure, combining various feature- and model- based discriminative training approaches, as well as a new feature-based domain adaptation technique, fMPE-MAP, which brought large improvement in the 2007 NIST Meeting Evaluation. The content of this talk includes contributions from my SRI colleague Dr. Andreas Stolcke, collaborators Dr. Mei-Yuh Hwang, Dr. Xin Lei at University of Washington, and Dr. Ozgur Cetin formerly of ICSI.

Bio:

Dr. Jing Zheng is a senior research engineer at Speech Technology and Research Laboratory, SRI International. He received his BS and PhD degrees from EE Department, Tsinghua University in 1994 and 1999. His recent research interests include automatic speech recognition, statistical machine translation, and speech-to-speech translation systems. He is one of the main contributors of SRI’s DECIPHER ? and DynaSpeak ? speech recognition technologies, and the creator of SRI’s SRInterp ? statistical machine translation technology. Currently he is the co task lead of SRI’s machine translation effort under Defense Advanced Research Projects Agency (DARPA)’s GALE program, teaming with eleven US and international partners, and the co PI of SRI’s  effort aiming to develop an Iraqi Arabic and English two-way speech-to-speech translation system for tactical use under DARPA’s TRANSTAC program. He has authored and coauthored more than 40 papers and articles in international conferences and journals. He has served as reviewer for IEEE Transaction on Audio, Speech and Language Processing, Computer Speech and Language, International Conference on Acoustics, Speech, and Signal Processing (ICASSP), International Conference of Pattern Recognition. He is a member of IEEE and ISCA.