<?xml version="1.0" standalone="yes"?>
<Paper uid="W97-0117">
  <Title>A Natural Language Correction Model for Continuous Speech Recognition 1</Title>
  <Section position="4" start_page="168" end_page="169" type="intro">
    <SectionTitle>
3. OUTLINE OF THE APPROACH
</SectionTitle>
    <Paragraph position="0"> The basic premise of our approach is to utilize linguistic and sublanguage-specifie information, even speaker-specific features, in order to improve the accuracy of speech transcription. For the experiments described in this paper, we considered the speech recognition system as a black box, that is, our efforts were directed at correcting the transcription errors in a post-processing mode,  rather than to improve the initial transcription. One advantage of this two-pass approach is its independence from any particular SRS, and indeed our Correction Box (C-Box) module can be used as a back-end of any speech system. Nonetheless, there are other possibilities as well, and we mention them briefly in the Future Directions section.</Paragraph>
    <Paragraph position="1"> The main goal of the C-Box approach is to generate a collection of context-sensitive text-based correction rules, in the form of xt,Y = xq~Y=, where X and Y are context word patterns, L is a word pattern in an erroneous transcription string, and R is a replacement string correcting transcription errors in L. In order to generate the correction rules we require both training and validation steps to optimize the C-Box performance with respect to a specific sublanguage as well as a speaker or a group of speakers. Therefore, the C-Box approach consists of the following sub-processes: (1) collecting training data, (2) aligning text samples, (3) generating correction rules, (4) validating correction rules, and (5) applying correction rules.</Paragraph>
  </Section>
class="xml-element"></Paper>