The description of iFlyTek Speech Lab system for NIST2009 Language Recognition Evaluation

Ying Xu, Yan Song, Yanhua Long, Hai-Bing Zhong, Li-Rong Dai

Abstract: In this paper, we present a description of the iFlyTek Speech Lab system for NIST 2009 LRE (Language Recognition Evaluation). The system consists of acoustic systems (i.e. GMM-MMI and GMM-SVM) and phonotactic systems (i.e. PPR 4-gram LM and PPR 3-gram SVM). First, we describe several state-of-the-art techniques applied in our language recognition system, such as FA (Factor Analysis), MMI (Maximum Mutual Information), and generative and discriminative LM (Language Modelling) techniques etc. Then, we will discuss our data preprocessing techniques for handling large amount training and development data, and the mismatch among different languages, genders and channels. Finally, the evaluation results for NIST2009's tasks and detailed analysis are given for 30, 10 and 3 seconds durations.