<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-2026">
  <Title>A STACK DECODER FOR CONTINOUS SPEECH RECOGNITION 1</Title>
  <Section position="11" start_page="197" end_page="197" type="concl">
    <SectionTitle>
FUTURE RESEARCH
</SectionTitle>
    <Paragraph position="0"> In this section, we sketch some ideas which, when implemented, are expected to improve the performance of the stack decoder.</Paragraph>
    <Paragraph position="1"> Installation of rapid match and language model. These imminent developments (described briefly above) are expected to boost the time and accuracy performance of the decoder significantly. Improvements in the partial transcription evaluator. In the section entitled &amp;quot;THE PARTIAL TRANSCRIPTION EVALUATOR&amp;quot; we have outlined PTE calculations that do not look ahead at speech data not 'consumed' by the hypothesized partial transcription. We are in the process of developing algorithms that do look ahead, and expect them (ion a theoretical basis) to be more effective than the ones we have implemented so far.</Paragraph>
    <Paragraph position="2"> The results on the 100-word task indicate that in the top choices, the longer words (which are typically content words) are generally recognized correctly, whereas uncertainty as to the correct shorter words (usually function words) exists. This suggests the following two improvements.</Paragraph>
    <Paragraph position="3"> Caching of word matches (terminology suggested by Doug Paul). In the present scheme, the scorers are repeatedly computing likelihoods of a word in the same time interval. Caching would save the results of such evaluations for future use.</Paragraph>
    <Paragraph position="4"> Special treatment of function words. One possibility is to treat the set of all function words (or even the set of all reasonable function word sequences) as a single model in the initial stages of recognition. As a post-processing step, the best choice for the function words will be determined. There are also certain benefits in this treatment as far as the language model is concerned, and research is currently being done on this aspect.</Paragraph>
  </Section>
class="xml-element"></Paper>