<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1079">
  <Title>Large Vocabulary Recognition of Wall Street Journal Sentences at Dragon Systems</Title>
  <Section position="6" start_page="390" end_page="390" type="evalu">
    <SectionTitle>
5. RESULTS ON WSJ DATA
</SectionTitle>
    <Paragraph position="0"> This section contains results on the 5000-word closed-vocabulary speaker-dependent verbalized punctuation version of the Wall Street Journal task, using the development test data. Table 1 lists results for all the WSJ speakers, displaying the word error rates using three different models. The first column contains the results of the first recognition run we did using models obtained by merely adapting our reference speaker's original models, using our old 8 parameter signal processing, yielding an overall word error rate of 35.6%. The second column contains our best 32 parameter unimodal models using the rePELing/respelling training strategy, after several iterations of training, with an overall error rate of 16.4%. Finally the last column contains the results of our first experiment recognizing Wall Street Journal sentences with the 32 stream tied mixture models described above, but based on only one segmentation step (segmentation into phonemes). This produced a word error rate of 14.8%.</Paragraph>
    <Paragraph position="1"> It is encouraging that the tied mixture models yielded better performance than did the unimodal models on 11 out of the 12 speakers, given that there has not yet been any opportunity for parameter optimization.</Paragraph>
  </Section>
class="xml-element"></Paper>