<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2005">
  <Title>Thai Grapheme-Based Speech Recognition</Title>
  <Section position="2" start_page="0" end_page="17" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Large vocabulary speech recognition systems traditionally use phonemes as sub-word units. This requires a pronunciation dictionary, which maps the orthographic representation of words into a sequence of phonemes. The generation of such a dictionary is both time consuming and expensive since it often requires linguistic knowledge of the target language. Several approaches to automatic dictionary generation have been introduced in the past with varying degrees of success (Besling, 1994; Black et al., 1998). Nevertheless, these methods still require post editing by a human expert or using another manually generated pronunciation dictionary.</Paragraph>
    <Paragraph position="1"> As a solution to this problem, grapheme-based speech recognition (GBSR) has been proposed recently (Kanthak and Ney, 2002). Here, instead of phonemes, graphemes - orthographic representation of a word - are used as the sub word units. This makes the generation of the pronunciation dictionary a trivial task. GBSR systems have been successfully applied to several European languages (Killer et al., 2003). However, because of the generally looser relation of graphemes to pronunciation than phonemes, the use of context dependent modeling techniques and the sharing of parameters across different models are of central importance.</Paragraph>
    <Paragraph position="2"> The variations in the pronunciation of phonemes in different contexts are usually handled by clustering the similar contexts together. In the traditional approach, decision trees are used to cluster polyphones - a phoneme in a specific context - together. Due to computational and memory constraints, individual trees are grown for each sub-state of each phoneme. This does not allow the sharing of parameters across polyphones with different center phonemes. Enhanced tree clustering (Yu and Schultz, 2003) lifts this constraint by growing trees which cover multiple phonemes.</Paragraph>
    <Paragraph position="3"> In this paper we present our experiments on applying grapheme-based speech recognition for Thai language. We compare the performance of the grapheme-based system with two phoneme-based systems, one using a hand-crafter dictionary, and the other using an automatically generated diction- null ary. In addition, we observe the effect of the enhanced tree clustering on the grapheme-based recognition system.</Paragraph>
  </Section>
class="xml-element"></Paper>