<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1038">
  <Title>Recognition Using Classification and Segmentation Scoring*</Title>
  <Section position="5" start_page="200" end_page="200" type="concl">
    <SectionTitle>
4. DISCUSSION
</SectionTitle>
    <Paragraph position="0"> In summary, we have described an alternative approach to speech recognition that combines classification and segmentation scoring to more effectively use segmental features. Our pilot experiments demonstrate that the classification-in-recognition approach can achieve performance comparable to the traditional formalism when frame-based features and equivalent Gaussian distributions are used, and that the segmentation score can be an important component of a classification approach. We anticipate performance gains with the additional use of segmental features in the classification component of the CIP~ model. We also plan to extend the model to incorporate context-dependent units.</Paragraph>
    <Paragraph position="1"> Our initial experiments with the segmentation probability indicate that finding this component via marginal probabilities computed with a detailed model may be more accurate than estimating boundary likelihood based on local observations, although this conclusion should be verified with experiments using a larger number of hypotheses per sentence than the 20 used so far.</Paragraph>
    <Paragraph position="2"> A number of improvements can be mode to both models, including using different choices for mixture components and eliminating some of the independence assumptions. Additionally, in the second method we plan to increase both the number of features per frame and the number of boundary-odjacent frames considered in computing the boundary probabilities. Eventually a hybrid method that combines elements of both approaches may prove to be the most effective.</Paragraph>
  </Section>
class="xml-element"></Paper>