<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-3002">
  <Title>The Importance of Prosodic Factors in Phoneme Modeling with Applications to Speech Recognition</Title>
  <Section position="7" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
5. Results
</SectionTitle>
    <Paragraph position="0"> For each prosodic variable (phrasal position or accent), tables were constructed listing the preferred tying of phonemes based on the log likehood results. Table 1, for example, lists all phonemes that should be tied on the basis of accent and those that should not. Similar tables exist for phrasal position and for the combination of both accent and phrasal position. Examples of certain phonemes are not present due to the relatively small size of the data set.</Paragraph>
    <Paragraph position="1"> Experimental results varied greatly between consonants and vowels. For consonants, there appeared to be an improvement in the model when phonemes are distinguished by phrasal position. Separation of accented and unaccented phrase initial consonants yielded no improvement to the model for most consonants. This implies that phrase initial accented and phrase initial unaccented phonemes should be merged into a single token. Accented consonants are also not benefited by positional information. Results indicate that phrase initial, medial and final accented phonemes can be merged together. Figure 6a illustrates a proposed model for the prosodic labeling of consonants based on these results.</Paragraph>
    <Paragraph position="2"> For vowels, a model showed improvement when the phoneme was separated into phrase initial, medial and final tokens. Vowel phoneme models also showed improvement when separated by accent. The accent on a vowel appears to be important regardless of phrasal position. These results suggest a six-way distinction should be used when modeling vowels and the proposed model is illustrated in figure 6b.</Paragraph>
  </Section>
class="xml-element"></Paper>