<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1016">
  <Title>On Using Written Language Training Data for Spoken Language Modeling</Title>
  <Section position="10" start_page="96" end_page="97" type="concl">
    <SectionTitle>
8. CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> We found the following interesting results: * Expanding the vocabulary with less frequent words does not substantially increase the word error on those words already in the vocabulary, but does eliminate many errors due to OOV words.</Paragraph>
    <Paragraph position="1"> * Doubling the amount of language model training text improves the language model, even though the text comes from different years than the test, and even though the text was not preprocessed into proper lexical forms.</Paragraph>
    <Paragraph position="2"> * It is possible to improve the quality of the language modeling text by modeling the differences between the  predicted rre~ding style and some examples of actual transcriptions.</Paragraph>
    <Paragraph position="3"> * Increasing the vocabulary size and language training hadL a bigger effect on spontaneous speech than it did for read speech.</Paragraph>
  </Section>
class="xml-element"></Paper>