File Information

File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/intro/92/h92-1081_intro.xml

Size: 2,039 bytes

Last Modified: 2025-10-06 14:05:18

<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1081">
  <Title>The Lincoln Large-Vocabulary HMM CSR*</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> The original Lincoln Tied-Mixture HMM CSR was implemented as a single-layer static network with a time-synchronous (TS) beam-pruned network-search strategy\[14\]. When used with a bigram language model (LM), this implementation generally requires an V ~ set of word interconnection links. This is tractable for vocabulary sizes on the order of 1K, but becomes intractable for vocabulary sizes (V) of 5K or more words. This implementation also is incompatible with or intractable for many forms of LM, such as recursive or trigram models. A stack decoder \[1,7,16\] with fast match\[2,3,5\] is used here to overcome the limitations of the original decoder structure.</Paragraph>
    <Paragraph position="1"> Previous work focused on a 1K word task, Resource Management (RM)\[18\], which could be handled adequately with the TS decoder. (The same decoder was also used on the ATIS task\[13\].) However, only the stack decoder was usable on the WSJ task. While the theory of the stack decoder is adequately established \[1,7,16\], many of the implementation details are still topics for research. One topic of particular interest is fast match techniques and structures. There are also a number of pragmatic issues to be resolved for stack decoders and *This work was sponsored by the Defense Advanced Research Projects Agency. The views expressed are those of the author and do not reflect the official policy or position of the U.S. Government. tree searches in general. (The stack decoder implements a tree search.) Once a functioning stack decoder was developed, it became possible to perform recognition experiments on the WSJ database. This allowed both further debugging and development of the stack decoder and exploration of acoustic modeling techniques.</Paragraph>
  </Section>
class="xml-element"></Paper>
Download Original XML