Using Tone Information in Thai Spelling Speech Recognition

Using Tone Information in Thai Spelling Speech Recognition Natthawut Kertkeidkachorn 1Department of Computer Engineering, Chulalongkorn University Bangkok, Thailand 2Department of Informatics The Graduate University for Advanced Studies, Tokyo, Japan Natthawut@nii.ac.jp Proadpran Punyabukkana Department of Computer Engineering, Chulalongkorn University Bangkok, Thailand Proadpran.p@chula.a c.th Atiwong Suchato Department of Computer Engineering, Chulalongkorn University Bangkok, Thailand Atiwong.s@chula.ac. th Abstract Spelling recognition is a workaround to recognize unfamiliar words, such as proper names or unregistered words in a dictionary, which typically cause ambiguous pronunciations. In the Thai spelling task, some alphabets cannot be differentiated by only spectral cues. In such cases, tonal cues play a critical role in distinguishing those alphabets. In this paper, we therefore introduce Thai spelling speech recognition, in which a tonal score, which represents a tonal cue, is adopted in order to re-rank N-best hypotheses of the first pass search of a speech recognition system. The Hidden Conditional Random Field (HCRF)-based Thai tone recognition, which was reported as the best approach for Thai tone recognition, is selected to provide tonal scores. Experimental results indicate that our approach provides the best error rate reduction of 23.85% from the baseline system, which is a conventional Hidden Markov Model (HMM)-based speech recognition system. Besides, another finding is that exploiting tonal scores in Thai spelling speech recognition could significantly reduce the ambiguity among some alphabets. 1 Introduction A spelling speech recognition system plays an important role in many kinds of applications, of which a domain contains unfamiliar words such as proper names. Since those words might not be pronounced straight-forwardly, an automatic speech recognition (ASR) system would have difficulty to recognize such words correctly. A practical efficient solution for handling such words in an ASR system is to pronounce them letter by letter. Nevertheless, in tonal languages, especially Thai, a spelling recognition task is a challenging task because merely consonantal sound and vowel sound cannot perfectly distinguish Thai alphabets. For example the “ข” alphabet and the “ค” alphabet are pronounced as \khƆ:\. Although the consonantal sound and the vowel sound of those alphabets are similar, their tones are significantly different. For the “ข” alphabet, its tonal sound is the rising tone, while the tonal sound of the “ค” alphabet is the mid tone. In Thai, tone information therefore not only expresses prosody as usual but also transmits explicit information, which characterizes lexical meanings of words (Luksaneeyanawin 1998). In this paper, we therefore introduce a Thai spelling speech recognition employing tonal scores, which can represent tonal information, in order to re-rank N-best hypotheses according to the first pass search of an ASR system. The Hidden Conditional Random Field (HCRF)-based Thai tone recognition, which had been reported as the state of the art for Thai tone recognition PACLIC 28 178 Copyright 2014 by Natthawut Kertkeidkachorn, Proadpran Punyabukkana, and Atiwong Suchato 28th Pacific Asia Conference on Language, Information and Computation pages 178–184 (Kertkeidkachorn et al. 2014), is selected to provide tonal scores. The rest of the paper is organized as follows. In Section 2, background knowledge on Thai …

Content truncated to 500 words