<?xml version="1.0" standalone="yes"?>
<Paper uid="J01-1002">
  <Title>Integrating Prosodic and Lexical Cues for Automatic Topic Segmentation</Title>
  <Section position="9" start_page="53" end_page="54" type="concl">
    <SectionTitle>
6. Conclusion
</SectionTitle>
    <Paragraph position="0"> We have presented a probabilistic approach to topic segmentation of speech, combining both lexical and prosodic cues. Topical word usage and lexical discourse cues are represented by language models embedded in an HMM. Prosodic discourse cues, such as pause durations and pitch resets, are modeled by a decision tree based on automatically extracted acoustic features and alignments. Lexical and prosodic features can be combined either in the HMM or in the decision tree framework.</Paragraph>
    <Paragraph position="1"> Our topic segmentation model was evaluated on broadcast news speech, and found to give competitive performance (around 14% error according to the weighted TDT2 segmentation cost metric). Notably, the segmentation accuracy of the prosodic  Tier, Hakkani-T(ir, Stolcke, and Shriberg Integrating Prosodic and Lexical Cues model alone is competitive with a word-based segmenter, and a combined prosodic/ lexical HMM achieves a substantial error reduction over the individual knowledge sources.</Paragraph>
  </Section>
class="xml-element"></Paper>