NIST 2007 Language Recognition Evaluation: From the Perspective of IIR

NIST 2007 Language Recognition Evaluation: From the Perspective of IIR* Haizhou Li, Bin Ma, Kong-Aik Lee, Khe-Chai Sim, Hanwu Sun, Rong Tong, Donglai Zhu, and Changhuai You Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR), Singapore {hli,mabin,kalee,kcsim,hwsun,tongrong,dzhu,echyou}@i2r.a-star.edu.sg Abstract. This paper describes the Institute for Infocomm Research (IIR) system for the 2007 Language Recognition Evaluation (LRE) conducted by the National Institute of Standards and Technology (NIST). The submitted system is a fusion of multiple state-of- the-art language classifiers using diversified discriminative language cues. We implemented several state-of-the-art algorithms using both phonotactic and acoustic features. We also investigated the system fusion and score calibration strategy to improve the performance of language recognition, and worked out a pseudo-key analysis approach to cross-validate the performance of the individual classifiers on the evaluation data. We achieve an equal-error-rate (EER) of 1.67 % on the close-set general language recognition test. Keywords: Automatic spoken language recognition, NIST Language Recognition Evaluation, phonotactic features, acoustic features, fusion system, pseudo key. 1. Introduction Automatic spoken language recognition (SLR) is a process of determining the identity of the language in a spoken document. As multilingual applications are demanded by the emerging need for globalization and the growing international business interflow, SLR has become an enabling technology in many applications such as multilingual conversational systems (Zue and Glass, 2000), multilingual speech recognition and translation (Waibel et al., 2000), and spoken document retrieval (Dai et al. 2003). It is also a topic of great importance in the areas of intelligence and security, where the language identities of recorded messages and archived materials need to be established before any information can be extracted. SLR technology also facilitates massive on-line language routing for voice surveillance over telephone network. The National Institute of Standards and Technology (NIST) has conducted a series of evaluations of SLR technology in 1996, 2003, 2005 and 2007 (NIST, 2007). The language recognition evaluations (LREs) focus on language and dialect detection in the context of conversational telephony speech. They are conducted to foster research progress, with the goals of exploring promising new ideas in language recognition, developing advanced technology incorporating these ideas, and measuring the performance of this technology. The Institute for * Copyright 2008 by Haizhou Li, Bin Ma, Kong-Aik Lee, Khe-Chai Sim, Hanwu Sun, Rong Tong, Donglai Zhu, and Changhuai You 46 22nd Pacific Asia Conference on Language, Information and Computation, pages 46–57 Infocomm Research (IIR) team has participated in the 2005 and 2007 NIST LREs and demonstrated the state-of-the-art technologies. One of the fundamental issues in SLR is to explore the discriminative cues for spoken languages. In the state-of-the-art language recognition systems, these cues mainly come from the acoustic features (Sugiyama, 1991; Torres-Carassquilo et al., 2002; Burget et al., 2006; Campbell et al., 2006) and phonotactic representations (Hazen and Zue, 1994; Zissman, 1996; Berkling and Barnard, 1994; Corredor-Ardoy et al., 1997; Li and Ma, 2005; Ma, Li, and Tong, 2007), which reflect different aspects of spoken language characteristics. Another issue is how to effectively organize and …

Content truncated to 500 words