<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1076">
  <Title>Session 13: CSR Search</Title>
  <Section position="3" start_page="0" end_page="386" type="intro">
    <SectionTitle>
2. Conclusions
</SectionTitle>
    <Paragraph position="0"> Some general conclusions can be made from the various attempts at improving the efficiency and accuracy of the search algorithms.</Paragraph>
    <Paragraph position="1"> First, while there are various tradeoffs that can be made related to pruning back the number of active hypotheses, or the size of the language model, etc, in general, these compromises quickly become damaging, in that they also increase the word error rate.</Paragraph>
    <Paragraph position="2"> The more effective approaches make use of two general techniques: shared computation, and multiple-pass strategies.</Paragraph>
    <Section position="1" start_page="0" end_page="385" type="sub_section">
      <SectionTitle>
2.1. Shared Computation
</SectionTitle>
      <Paragraph position="0"> Two effective ways to share computation are to use tree structures, and to perform bottom-up processing.</Paragraph>
      <Paragraph position="1"> Tree structures, both at the phonetic and language modeling  level reduce computation by a large factor since the computation for the initial portions of similar words can be shared. By the time the computation gets to the ends of the words, most of the words have been eliminated.</Paragraph>
      <Paragraph position="2"> Bottom-up processing means that a system examines the input without regard to the surrounding context, and uses these scores in ,various combinations depending on the global context. Thus, the repeated scoring of the same acoustic events in different language model contexts is avoided.</Paragraph>
    </Section>
    <Section position="2" start_page="385" end_page="386" type="sub_section">
      <SectionTitle>
2.2. Multiple-Pass Strategies
</SectionTitle>
      <Paragraph position="0"> There are several multi-pass search strategies that have found beneficial use when real-time is desired. The problem is that, even though it would be nice to use all of the knowledge sources at once to obtain their full integration, this is just too expensive for the size of problems we are trying to handle, and the currently available hardware. The single-pass search employed by Cambridge University certainly showed that there is much to be gained from efficient sharing, primarily through the use of dynamically compiled tree structures.</Paragraph>
      <Paragraph position="1"> However, at the current time, it seems unlikely that this approach could be pushed all the way to real-time processing.</Paragraph>
      <Paragraph position="2"> The multiple-puss strategies discussed here include fast match algorithms, using vector quantization as an approximation to eliminate most of the computation for Gaussians, use of N-best searches with reduced models followed by rescoring with more detailed models, and use of lattices in much the same way. In addition, the use of the forward-backward search technique allows the later passes to make more effective use of the pruning information derived from earlier passes.</Paragraph>
      <Paragraph position="3"> The multiple-pass search strategies often can save several orders of magnitude in search computation, thus making real-time conceivable.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>