Statistical Method of Building Dialect Language Models for ASR Systems

Proceedings of COLING 2012: Technical Papers, pages 1179–1194, COLING 2012, Mumbai, December 2012. Statistical Method of Building Dialect Language Models for ASR Systems Naoki HIRAYAMA1 Shinsuke MORI1,2 Hiroshi G. OKUNO1 (1) Graduate School of Informatics, Kyoto University (2) Academic Center for Computing and Media Studies, Kyoto University Yoshida-honmachi, Sakyo-ku, Kyoto, Japan hirayama@kuis.kyoto-u.ac.jp, forest@i.kyoto-u.ac.jp, okuno@i.kyoto-u.ac.jp ABSTRACT This paper develops a new statistical method of building language models (LMs) of Japanese dialects for automatic speech recognition (ASR). One possible application is to recognize a variety of utterances in our daily lives. The most crucial problem in training language models for dialects is the shortage of linguistic corpora in dialects. Our solution is to transform linguistic corpora into dialects at a level of pronunciations of words. We develop phoneme- sequence transducers based on weighted finite-state transducers (WFSTs). Each word in common language (CL) corpora is automatically labelled as dialect word pronunciations. For example, anta (Kansai dialect) is labelled anata (the most common representation of ‘you’ in Japanese). Phoneme-sequence transducers are trained from parallel corpora of a dialect and CL. We evaluate the word recognition accuracy of our ASR system. Our method outperforms the ASR system with LMs trained from untransformed corpora in written language by 9.9 points. KEYWORDS: spoken language, dialect, language model, weighted finite-state transducer (WFST). 1179 1 Introduction Automatic speech recognition (ASR) systems for spoken language are yet far from practical use. ASR systems for written sentences have been widely studied, and recognition accuracy has rapidly improved. In contrast, recognition accuracy is drastically lower for spontaneous speech (Anusuya and Katti, 2009, p. 194). People in their daily lives do not actually speak in a stable way like written sentences. Their speeches include casual expressions, fillers, and even vocabulary specific to dialects. This paper especially handles improving Japanese dialect ASR. Most previous application systems with speech interface have assumed well-formed sentences in the common language (CL), although they have assumed non-expert speakers. Non-expert speakers will obviously utter informal expressions other than those in written language, and even words specific to their own dialect; dialect ASR systems have difficulty in recognition accuracy or scalability. Dialects in the world have various kinds of differences (Benincà, 1989). The major differences between dialects and the CL are categorized into the following types: 1) pronunciation, 2) vocabulary, and 3) word order. The first type belongs to the difference in acoustic features, while the second and third belong to that in linguistic ones. Canadian English contains all of three types; 1) /tu/ is pronounced as /tju/, 2) ‘high school’ is called ‘collegiate institute’, and 3) ‘next Tuesday’ is changed into ‘Tuesday next’ (Woods, 1979). Many North American varieties of French have a tendency to take SVO (subject-verb-object) word order (Gadet and Jones, 2008). In Japanese dialects, the difference of vocabulary is characteristic, e.g., tabe n is used instead of tabe nai (do not eat) (Gottlieb, 2005). Our method in this paper focuses on differences in pronunciation of vocabulary between dialects, which correspond to the first and second types. …

Content truncated to 500 words