<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-1613">
  <Title>A Study of Automatic Pitch Tracker Doubling/Halving &amp;quot;Errors&amp;quot;</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> A task-oriented conversational speech database was manually annotated for pitch, but work remains to make the database precise enough for intonation research. This work focussed on potential halving and doubling errors of pitch trackers, and on evaluation of pitch trackers with respect to a final boundary tone classification.</Paragraph>
    <Paragraph position="1"> Statistically significantly better classification results were achieved with manual verification of pitch data based on wideband spectrograms and speech waveform information. These results were achieved even though the hand measurements appeared to be very similar to the automatic measurements. Based on the very preliminary results of this study, the following two conclusions can be made at this time: one, that automatic pitch measurements still might not be as accurate as needed in order to make generalizations about intonation contours in conversational speech; and secondly, that the combination of a wideband spectrogram and signal shape is a useful starting point for creating large-scale hand-verified pitch tracks of conversational speech.</Paragraph>
  </Section>
class="xml-element"></Paper>