<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-2033">
  <Title>Improved HMM Models for High Performance Speech Recognition</Title>
  <Section position="6" start_page="253" end_page="254" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> We draw several conclusions from this work: * Supervising the VQ with phoneme identity does not help overall recognition performance.</Paragraph>
    <Paragraph position="1"> Shared mixtures in the decoder reduces error rate by 10%-20% depending on the grammar, but after smoothing only by 5%-20%.</Paragraph>
    <Paragraph position="2"> We found no improvement for replacing the heuristically derived weights for the context-dependent models with weights determined by deleted estimation. null We have implemented an algorithm for MMI training in continuous speech that uses alternatives generated by the N-Best algorithm. Initial experiments to optimize the three feature set weights using this procedure reduced word error rate by 10%.</Paragraph>
    <Paragraph position="3">  As expectecL using cross-word tfiphone models reduced word error rate by 30%.</Paragraph>
    <Paragraph position="4"> The word error rate using the Word-Pair grammar is now close to 2%, depending on the test set. When no grammar is used the error rate was 10.6% on the Oct. '89 test set. Due to the very low error rate with the Word-Pair grammar, we will use the statistical class grammar (Derr, 1989) for most of our testing as it will be easier to measure improvements using this more difficult and more realistic grammar.</Paragraph>
  </Section>
class="xml-element"></Paper>