<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1102">
  <Title>Robust Continuous Speech Recognition Technology Program Summary *</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
RECENT ACCOMPLISHMENTS
</SectionTitle>
    <Paragraph position="0"> Developed and improved the Lincoln large-vocabulary tied-mixture HMM CSR, including stack decoder search, acoustic fast-match, and cross-word and sex-dependent acoustic models, and applied this CSR in the November 1993 evaluation tests; the new system showed a 42 percent improvement in error rate compared to the November 1992 evaluation test system.</Paragraph>
    <Paragraph position="1"> Developed and successfully tested recognition-time adaptation techniques for large-vocabulary CSR in the November 1993 evaluation tests.</Paragraph>
    <Paragraph position="2"> Developed tests on data-driven allophonic tree clustering smoothing techniques for best use of available training data. Developed Bayesian smoothing techniques for triphones and obtained promising initial results on CSR corpora.</Paragraph>
    <Paragraph position="3"> Continued contributions to ARPA CSR corpus development and evaluation, including contribution of stochastic language models to all sites for the 1993 evaluation tests; provided the ARPA CSR community with text processing software tools for large-vocabulary corpus development.</Paragraph>
    <Paragraph position="4"> Organized and chaired the ARPA Spoken Language Technology and Applications Day (SLTA 93), which has produced very promising results in catalyzing technology transition of spoken language technology into military and civilian applications. null</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="459" type="metho">
    <SectionTitle>
PLANS
</SectionTitle>
    <Paragraph position="0"> Continue to develop large-vocabulary stack decoder-based HMM CSR, with particular focus on improvement of acoustic fast-match techniques, Develop advanced acoustic modelling techniques including data-driven decision-tree-based triphone smoothing.</Paragraph>
    <Paragraph position="1"> Develop run-time adaptation techniques for both acoustic HMM parameters and for stochastic language model parameters; include adaptation to speaker, channel, environment,, and task.</Paragraph>
    <Paragraph position="2"> Continue to define and develop spoken language technology applications, with particular focus on recognition and understanding of spoken messages in a command and control environment; also continue follow-up on other application opportunities produced by SLTA 93.</Paragraph>
    <Paragraph position="3"> Chair the 1994 ARPA Human Language Technology Workshop. null</Paragraph>
  </Section>
class="xml-element"></Paper>