<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1066">
  <Title>SIGNAL PROCESSING FOR ROBUST SPEECH RECOGNITION</Title>
  <Section position="10" start_page="334" end_page="334" type="concl">
    <SectionTitle>
5. SUMMARY AND CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> In this paper we describe a number of procedures that improve the recognition accuracy of the SPHINX-II system in unknown acoustical environments. We found that the use of MFCDFN and phone-dependent cepstral normalization reduces the error rate by 40 percent compared to that obtained with CMN alone. The use of Baum-Welch codebook adaptation with MFCDCN reduces the error rate by 37 percent compared to that obtained with CMN alone. The use of reduced-frequency processing reduces error rates for telephone-bandwidth speech by 58 percent compared to the rate observed for conventional signal processing. The performance of these systems for the 1993 CSR Spoke 5 and Spoke 8 evaluatious is described.</Paragraph>
  </Section>
class="xml-element"></Paper>