<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-2034">
  <Title>Speaker Adaptation Using Multiple Reference Speakers</Title>
  <Section position="5" start_page="260" end_page="261" type="concl">
    <SectionTitle>
4 Summary
</SectionTitle>
    <Paragraph position="0"> We have described our speaker-adaptation system in terms of the two speaker-transformations used to make one speaker look like another; speech normalization and PDF mapping. Experimental results indicate that the speech normalization can be improved by feature conditioning, whereas the PDF mapping is relatively insensifive to improvements in the normalization. Also we have shown that the choice of any single reference speaker is not an important issue, indicating that improvements to the reference model are likely to be gained only by using multiple reference speakers.</Paragraph>
    <Paragraph position="1"> We have reported baseline system (single reference speaker) test results of 7.4% word error rate for the word-pair grammar and 28.7% for no grammar on the designated Oct. '89 DARPA evaluation test set. This performance is comparable to the best speaker-independent results being reported today, but with considerably less ~  effort required to collect the reference training material (1 speaker vs. 100, 600 utterances vs. 4000).</Paragraph>
    <Paragraph position="2"> We have proposed a new method of utilizing speech from multiple reference speakers by transforming them to a single common feature space before pooling. Preliminary experiments have shown a five-fold reduction in error rate for using the proposed normalization on a 12 speaker pooled model compared to a single speaker model. We propose to test our approach on the speaker-independent portion of the DARPA Resource Management database in the near future.</Paragraph>
  </Section>
class="xml-element"></Paper>