<?xml version="1.0" standalone="yes"?>
<Paper uid="H91-1011">
  <Title>Modelling Context Dependency in Acoustic-Phonetic and Lexical Representations 1</Title>
  <Section position="5" start_page="74" end_page="74" type="evalu">
    <SectionTitle>
BENCHMARK RESULTS
</SectionTitle>
    <Paragraph position="0"> In connection with the Fourth DARPA Speech and Natural Language workshop, we participated in the benchmark evaluation of the SUMMIT system on the Resource Managenemet task, using the February-91 test set released by NIST. The system used context-normalized input dimensions with mixture Gaussian models, and was trained on the standard 109-speaker training set. The results are shown in Table 2.</Paragraph>
    <Paragraph position="1"> Comparing the last row of Table 1 with Table 2, we see that our system's performance is quite similar on the two different test sets. We are encouraged by the results of our first attempt at context-dependent modelling. We expect that additional performance gain can be realized when more complex models are introduced.</Paragraph>
    <Paragraph position="2">  agement task with a perplexity 60 language model, using the February-91 test set released by NIST. The system used context-normalization trees for each input dimension, with mixture Gaussian models.</Paragraph>
  </Section>
class="xml-element"></Paper>