<?xml version="1.0" standalone="yes"?>
<Paper uid="J00-2003">
  <Title>A Multistrategy Approach to Improving Pronunciation by Analogy</Title>
  <Section position="9" start_page="213" end_page="216" type="concl">
    <SectionTitle>
8. Conclusions
</SectionTitle>
    <Paragraph position="0"> We have extended previous work in pronunciation by analogy (PbA) principally by experimenting with multiple strategies for producing pronunciations. The combina- null Results of letter-to-stress conversion for the 31 possible combinations of scoring strategy using the product rule. Rank(C) is the rank of the result according to the number of strategies (in the range 1 to 5) included in the final score. Rank(W) is the rank of the result according to word accuracy. The Spearman rank correlation coefficient rs is 0.7411, which is very highly significant.</Paragraph>
    <Paragraph position="1">  tion of different scoring strategies for the shortest paths in the pronunciation lattice has been shown to deliver statistically significant improvements: a best result of 65.5% words correct has been obtained for letter-to-phoneme conversion using approximately 20,000 manually aligned words of Webster's Pocket Dictionary. This compares with a figure of 63.0% for the best-performing single-scoring strategy (frequency of the same pronunciation, FSP) and 61.7% for our best preliminary result. Examination of the pronunciation lattices, however, reveals that the upper bound on performance of a method based on selecting among shortest paths is 85.1%, so that there is scope for further improvement yet.</Paragraph>
    <Paragraph position="2"> The way of combining scores is simply by summation or multiplication of a points score awarded on the basis of a pronunciation's rank. The product rule of combination is found to perform only very marginally better than the sum rule: the difference is not significant. We have not yet expended much effort in determining exactly which scores should be used. To this end, a preliminary analysis of errors has been done. This  reveals that (in common with other approaches to letter-to-sound conversion, such as linguistic rules), the translation of vowel letters is especially problematic. Future work should therefore attempt to find scoring methods and combination techniques that deal effectively with the vowels.</Paragraph>
    <Paragraph position="3"> We have also studied the related problems of mapping phonemes to letters and letters to lexical stress, which are also important in speech technology. Our results show that the former problem is easier than letter-to-phoneme conversion while the latter is harder--at least, when attacked by the method of analogy. Once again, however, the multistrategy approach has the potential to deliver significant performance gains.</Paragraph>
  </Section>
class="xml-element"></Paper>