Thai Spelling Recognition Using a Continuous Speech Corpus

Thai Spelling Recognition Using a Continuous Speech Corpus Chutima Pisarn Sirindhorn International Institute of Technology 131 Moo 5 Tiwanont Rd., Bangkadi, Muang, Phathumthani, Thailand, 12000 chutimap@siit.tu.ac.th Thanaruk Theeramunkong Sirindhorn International Institute of Technology 131 Moo 5 Tiwanont Rd., Bangkadi, Muang, Phathumthani, Thailand, 12000 thanaruk@siit.tu.ac.th Abstract Spelling recognition is an approach to enhance a speech recognizer’s ability to cope with incorrectly recognized words and out-of-vocabulary words. This paper presents a general framework for Thai speech recognition enhanced with spelling recognition. In order to implement Thai spelling recognition, Thai alphabets and their spelling methods are analyzed. Based on hidden Markov models, we propose a method to construct a Thai spelling recognition system by using an existing continuous speech corpus. To compensate the difference between spelling utterances and continuous speech utterances, the adjustment of utterance speed is taken into account. Assigning different numbers of states for syllables with different durations is helpful to improve the recognition accuracy. Our system achieves up to 79.38% accuracy. 1 Introduction Nowadays, several works on automatic speech recognition (ASR) for continuous speech are being developed, not only system that rely on dictionary, but also the recognition on out of vocabulary circumstances. In a situation of misrecognition and out-of-vocabulary words, a practical and efficient solution that would assist the ASR is to equip the system with a spelling recognition subsystem, where users can spell out a word letter by letter. Spelling recognition is a challenging task with a high interest for directory assistance sevices, or other applications where a large number of proper names or addresses are handled. Many works that focus on spelling recognition were widely developed in several languages, for instance, English, Spanish, Portuguese and German. In (San- Segundo et al., 2001) the hypothesis-verification Spanish continuous spelled proper name recognition over the telephone was proposed. In this work, several feature sets were investigated by using models of neural networks. In their succeeding work (San-Segundo et al., 2002), three different recognition architectures, including the two-level architecture, the integrated architecture and the hypothesis-verification architecture, are analyzed and compared. In (Rodrigues et al., 1997), a Portuguese speaker -independent system for recognizing an isolated letter was introduced. The system dealt with speech utterances over a telephone line using Hidden Markov Model (HMM). A number of experiments were made over four different perplexity language models. Mitchell and Setlur (1999) proposed a fast list matcher to select a name from a name list that was created from an n-best letter recognizer on spelling over a telephone line recognition task. In (Bauer and Junkawitsch, 1999), an approach is proposed to combine word recognition with spelling recognition in a user-friendly manner as a fall back strategy. As a German city name recognizer, the system was applied to directory assistance services. Unlike other languages, spelling in Thai has several styles. One of them is similar to spelling in English, i.e., /d-ii//z-oo//g-ii/ for “dog”. There are three more methods in Thai spelling, where some syllables are inserted to make it clearer for the hearer. One is to …

Content truncated to 500 words