<?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="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2. The Database
</SectionTitle>
    <Paragraph position="0"> Boston University's Radio News Corpus (1995) was used for all experiments. The speakers from this corpus that were analyzed were F1A, F2B, and M2B. The usable data from these three speakers consisted of 259 phn : phrase medial phn! : phrase medial, accented  represents some generic phoneme.</Paragraph>
    <Paragraph position="1"> wav files containing 18270 words. All the wav files that were used were accompanied by two types of prosodic transcription files, .brk and .ton files. The corpus was labeled according to the ToBI standard. Silverman et al (1992) explain the labeling system in detail. It will not be described in this paper. The .brk files specify a ToBI break index (0-4) for every spoken word in the associated wav file. For the experiments, the only boundary distinguished was the intonational phrase boundary (ToBI index 4). All other boundary types (indices 0-3) were grouped together. There were 3855 intonational phrase boundaries in the data set.</Paragraph>
    <Paragraph position="2"> The .ton files label the times in which an accented vowel occurs. The most abundant accent label was H* which occurs in a ratio of about 10 H* for every single L*. Other accent types do occur, but most include H* in bitonal accent.</Paragraph>
  </Section>
class="xml-element"></Paper>