<?xml version="1.0" standalone="yes"?>
<Paper uid="H91-1056">
  <Title>LEXICAL ACCESS WITH A STATISTICALLY-DERIVED PHONETIC NETWORK</Title>
  <Section position="10" start_page="291" end_page="291" type="concl">
    <SectionTitle>
7. RESULTS
</SectionTitle>
    <Paragraph position="0"> At this time we have a simple version of the model described here running. We have not yet implemented the wordcoartlculation component and the lexical likelihood model has the form: P(y, dlw ) ~ P(ylw)P(dIw) (7.1).</Paragraph>
    <Paragraph position="1"> In other words, the duration model does not include the phone sequence only the phoneme sequence (cf. Eq. 2.2).</Paragraph>
    <Paragraph position="2"> Testing the model on the February '89 DARPA resource management test set and using the word-pair grammar, we achieved 85.7% word correct and 83.2% word accurary. Word insertion were 2.4% and deletions were 3.5%. This is with a phone recognizer that is achieves an estimated 81.5% phone correct and 76.0% phone accuracy on the same test set, using automatically derived phonetic transcriptions \[see 1\].</Paragraph>
    <Paragraph position="3"> We are encouraged by this since it is a considerable improvement over this system's progenitor and approaching the best results reported for phone-based recognition. This improvement is due both to much better phone recognition and to improved lexical access with this approach.</Paragraph>
    <Paragraph position="4"> We believe considerable further improvement will come when we include better duration information, word co-articulation, and, most importantly, when we input a phone lattice with recognizer alternatives rather than just the best guess. We have, in fact, implemented a crude version of a lattice in which the segmentation produced by the best guess is used, but alternative phones and their likelihoods are included. This performed 88.5% phone correct and 87.2% phone accuracy on the the Feb '89 test set. We are now implementing a structure that allows a true lattice that will allow alternative segmentations.</Paragraph>
  </Section>
class="xml-element"></Paper>