<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-1062">
  <Title>Unlimited vocabulary speech recognition for agglutinative languages</Title>
  <Section position="7" start_page="492" end_page="492" type="concl">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> This work presents statistical language models trained on different agglutinative languages utilizing a lexicon based on the recently proposed unsupervised statistical morphs. To our knowledge this is the first work in which similarly developed subword unit lexica are developed and successfully evaluated in three different LVCSR systems in different languages. In each case the morph-based approach constantly shows a significant improvement over a conventional word-based LVCSR language models. Future work will be the further development of also the grammatical morph-based language models and comparison of that to the current approach, as well as extending this evaluation work to new languages.</Paragraph>
  </Section>
class="xml-element"></Paper>