<?xml version="1.0" standalone="yes"?>
<Paper uid="H91-1012">
  <Title>BYBLOS SPEECH RECOGNITION BENCHMARK RESULTS</Title>
  <Section position="11" start_page="81" end_page="81" type="concl">
    <SectionTitle>
CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> We have reported several new benchmark speech recognition results for both the RM corpus and the new ATIS corpus. On RM, using the standard 109 speaker training set and the word-pair grammar, the word error rate for the BYBLOS system was 3.8%. Surprisingly our new SI paradigm, using only 12 training speakers, achieved the same resnitl In addition, we have demonstrated that SI performance is generally very bad for speakers with s-~ong dialects. But we have achieved a 5-fold reduction in error rate for these speakers by using speaker adaptation from only 40 training utterances.</Paragraph>
    <Paragraph position="1"> For the ATIS corpus we developed several new techniques based on the N-Best paradigm. These have allowed us to use cross-word triphone models and a 4-gram statistical grammar efficiently in the recognition. We have improved performance over a baseline condition by 20%-30% by using additional training, models of nonspeech, and a 4-gram class grammar. Our preliminary conclusion is that most of this gain is due to the smoothing of the grammar by classes. The spontaneous speech effects that appear in this corpus clearly present a new set of difficult problems, since the error rates are about 4 times higher than for the RM corpus.</Paragraph>
  </Section>
class="xml-element"></Paper>