<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-2044">
  <Title>ACOUSTICAL PRE-PROCESSING FOR ROBUST SPEECH RECOGNITION</Title>
  <Section position="11" start_page="315" end_page="315" type="evalu">
    <SectionTitle>
EXPERIMENTAL RESULTS
</SectionTitle>
    <Paragraph position="0"> Figure 2 summarizes the experimental results obtained using the Alphanumeric database when the system was trained and tested on the two types of microphones, in either the baseline conditions, or with spectral normalization and spectral subtraction. In each of the two panels, the word accuracies obtained for the two baseline conditions when the system was trained and tested using the same microphones are indicated by the horizontal dotted lines. It can be seen that in each case, the use of spectral normalization and subtraction provides increasing improvement to the recognition accuracy obtained in the &amp;quot;cross&amp;quot; conditions, without almost no degradation of the recognition accuracy observed when the system is trained and tested using the same microphone. In fact, the recognition accuracy obtained with spectral subtraction in the &amp;quot;cross&amp;quot; conditions approaches that obtained when the system is trained on the same microphone that it is tested on. On the other hand, we have not yet been able to significantly improve the performance of the system when it is trained and tested on the Crown PZM microphone. We briefly describe some of the strategies we are presently considering toward that end.</Paragraph>
  </Section>
class="xml-element"></Paper>