<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1014">
  <Title>Word Error Due To: Modeling</Title>
  <Section position="6" start_page="76" end_page="76" type="concl">
    <SectionTitle>
4. SUMMARY
</SectionTitle>
    <Paragraph position="0"> We have shown superior speech recognition performance with only a modest amount of training speech by aggressively handling the idiosyncrasies of this corpus. All utterances that are degraded due to severe disfluencies or problems with data-capture are eliminated from the training set.</Paragraph>
    <Paragraph position="1"> The excessively long and numerous segments of ambient noise in the data are removed from consideration by a good speech detector in the front-end. The very numerous hesitation phenomena are automatically located and then explicitly modeled where they occur in the training. Nonspeech events, such as filled-pauses, are made very unlikely in the grammar to clamp the false alarm rate.</Paragraph>
    <Paragraph position="2"> In addition, the trigram language model on word classes significantly improved recognition performance compared to a bigram model.</Paragraph>
    <Paragraph position="3"> With these improvements, the official BYBLOS speech recognition results for the February '92 DARPA evaluation were 6.2% word error for the Class A+D subset of the test and 9.4% overall. Both of these results were significantly better than any other speech system tested.</Paragraph>
    <Paragraph position="4"> Finally, we have shown how the N-best interface between the speech and natural components reduces the error rate compared to considering the top choice only. This was shown to be true whether a robust fragment processor was used as a fall-back or not.</Paragraph>
    <Paragraph position="5"> The official SLS result for HARC was a weighted error of 43.7. This was the best overall result for a spoken language system in the February '92 DARPA evaluation.</Paragraph>
  </Section>
class="xml-element"></Paper>