<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2005">
  <Title>Thai Grapheme-Based Speech Recognition</Title>
  <Section position="6" start_page="19" end_page="19" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper we presented the results for applying grapheme-based speech recognition to Thai language. We experimented with different settings for the initial context independent system, different number of acoustic models and different contexts for the polygraphemes. We also tried the enhanced tree clustering method as a means of sharing parameters across models. The results show that the system with 500 acoustic models based on trigraphemes produce the best results. Additionally, the enhanced tree clustering significantly improves the recognition accuracy of the grapheme-based system. Our system outperformed a phoneme-based system that uses an automatically generated dictionary. These results are very promising since they show that the grapheme-based approach can be successfully used to generate speech recognition systems for new languages using little linguistic knowledge.</Paragraph>
  </Section>
class="xml-element"></Paper>