<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1080">
  <Title>A One Pass Decoder Design For Large Vocabulary Recognition</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1. INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> Hidden Markov Models (HMMs) have been used successfully in a wide variety of recognition tasks ranging from small isolated word systems assisted by heavily constrained grammars to very large vocabulary unconstrained continuous speech systems. Part of the success of HMMs is due to the existence of computationally efficient algorithms for both the training of the models (the Baum-Welch algorithm) and for the decoding of unknown utterances (the Viterbi algorithm). However, as recognition tasks have become more complex, the decoding process has become more difficult due to the increasing size of network needed in a conventional Viterbi decoder. In particular, using cross-word triphones or long span language models can lead to order of magnitude increases in the size of a static network.</Paragraph>
    <Paragraph position="1"> A variety of schemes have been proposed to reduce the computation required for recognition \[2,7\]. Most make use of fast-matches or preliminary passes using simplified acoustic or linguistic models to constrain the search space for a final pass that uses the most detailed and accurate models available. Unfortunately, these preliminary matches can introduce errors that subsequent passes are unable to correct. If the first pass could use the best acoustic and language models available it would allow greater constraints to be placed on the search space without increasing the error rate. This paper describes a scheme which allows this through the use of a dynamically built tree-structured network. This approach avoids the compromises inherent in using fast-matches or preliminary passes and is both simple and efficient to implement.</Paragraph>
    <Paragraph position="2"> The remainder of this paper is organised as follows. Section 2 discusses the main features of conventional time-synchronous Viterbi decoding and some of the ways in which it can be improved. In section 3, a one-pass decoder that implements a beam-pruned Viterbi search through a tree-structured dynamic network is then described. Section 4 presents some experimental results on the Wall Street Journal task and, finally, section 5 presents our conclusions on this work.</Paragraph>
  </Section>
class="xml-element"></Paper>