<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1083">
  <Title>apos;Fable 5: Evaluation Male Speakers with Extra Training Data Speaker Baseline Larger-Set Training Training</Title>
  <Section position="11" start_page="426" end_page="426" type="concl">
    <SectionTitle>
9. SUMMARY
</SectionTitle>
    <Paragraph position="0"> This is a preliminary report demonstrating that the DECIPHER TM speech recognition system was ported from a 1,000-word task (ATIS) to a large vocabulary (5,000-word) task (DARPA's CSR task). We have achieved word error rates between of 16.6% and 17.1% as measured by NIST on DARPA's February 1992 Dry-Run WSJ0 evaluation where no test words were outside the prescribed vocabulary. We evaluated using alternate microphone data and found that the error rate increased only by 62%. Finally, by increasing the amount of training data, we were able to achieve an error rate that matched the error rates reported for this task from 600 sentence/speaker speaker-dependent systems.</Paragraph>
    <Paragraph position="1"> This could not have been done without substantial support from the rest of the DARPA community in the form of speech data, pronunciation tables, and language models.</Paragraph>
  </Section>
class="xml-element"></Paper>