<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-2044">
  <Title>ACOUSTICAL PRE-PROCESSING FOR ROBUST SPEECH RECOGNITION</Title>
  <Section position="16" start_page="316" end_page="317" type="concl">
    <SectionTitle>
SUMMARY
</SectionTitle>
    <Paragraph position="0"> We found that the use of desk-top microphones like the Crown PZM increase the error rate by allowing weak phonetic events to become confused with silences and vice-versa. The spectral subtraction and normalization routines we developed provide considerable improvement in recognition accuracy when the system is tested using a different microphone from the one it was trained on, but further work must be done to improve the absolute level of performance obtained when Sphinx is trained and tested using the Crown PZM.</Paragraph>
  </Section>
class="xml-element"></Paper>