<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-3007">
  <Title>Word Fragment Identification Using Acoustic-Prosodic Features in</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
4 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> Word fragment detection is very important for identifying disfluencies and improving speech recognition. In this paper, we have investigated the problem of word fragment detection from a new approach. We extracted a variety of prosodic features and voice quality measurement to capture the possible acoustic cues at the loca-tion of word fragments. Experimental results show that acoustic-prosodic features provide useful information for word fragment detection. These results offer an alternative view of the approach from building acoustic models in a recognizer to handle word fragments and suggest that speech analysis can be quite relevant to building better speech recognition approaches.</Paragraph>
    <Paragraph position="1"> These results are very preliminary. For example, experiments were only conducted using the downsampled data due to the extremely highly skewed data distribution.</Paragraph>
    <Paragraph position="2"> The current word fragment detection method would generate many false alarms in the real test situation, i.e. nondownsampled data. In addition, large corpora must certainly be examined and more sophisticated versions of the measures than we have used should be investigated, especially the voice quality measurements we used. However, as a first approximation of the characterization of word fragments via the acoustic-prosodic cues, we find these results encouraging. In particular, our ability to identify word fragments using only a few features seems promising. The potential features revealed by the experiments in this paper may be helpful to the method of building acoustic model for word fragment detection. Furthermore, we also need to investigate the performance when applying such an approach to the speech recognition results. Finally, a unified framework for word fragment and the disfluency detection is also a future direction of our work.</Paragraph>
  </Section>
class="xml-element"></Paper>