<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1055">
  <Title>Phonological Parsing for Bi-directional Letter-to-Sound/Sound-to-Letter Generation 1</Title>
  <Section position="8" start_page="293" end_page="293" type="ackno">
    <SectionTitle>
DISCUSSION
</SectionTitle>
    <Paragraph position="0"> Our current work demonstrates the utility of a hierarchical framework, which is relatively rich in linguistic knowledge, for bi-directional letter-to-sound/soundto-letter generation. The use of layered bigrams in our hierarchy is extendable to encompass natural language constraints \[10\], prosody, discourse and perhaps even dialogue modeling constraints on top, as well as phonetics and acoustics at the bottom. As such this paradigm should be particularly useful for applications in speech synthesis, recognition and understanding.</Paragraph>
    <Paragraph position="1"> The versatility of our framework can lead to a variety of applications. These range from a lexical representation for large-vocabulary recognition, which provides semantic information, syntactic information, and a clustering mechanism for fast match \[12\], to a low-perplexity langnage model for character recogition tasks, where our system gives a test set perplexity of 8.0 as constrasted Swe did not investigate the cases where maximum word fragment lengths are set beyond 6, due to computational limitations, and the vast number of training parameters required.</Paragraph>
    <Paragraph position="2"> with 11.3 from a standard letter bigram. In the near future, we plan to report on our robust parsing mechanism for extending coverage, and to experiment with alternative search strategies in the layered bigrams framework.</Paragraph>
  </Section>
class="xml-element"></Paper>