<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-1011">
  <Title>Speaker Adaptation from Limited Training in the BBN BYBLOS Speech Recognition System</Title>
  <Section position="6" start_page="104" end_page="104" type="concl">
    <SectionTitle>
5. Conclusion
</SectionTitle>
    <Paragraph position="0"> Three improvements to the DTW-based speaker adaptation method have been combined to achieve a 45% overall reduction in recognition word error rate on development test data. The largest single improvement was due to the addition of a codebook derived from a set of cepstral derivative features. This improvement does not affect the estimation of the between-speaker transformation.</Paragraph>
    <Paragraph position="1"> This suggests that further improvements to the speaker-dependent prototype model can lead to significant improvements in the adapted model's performance.</Paragraph>
    <Paragraph position="2"> The performance of the system on new evaluation test data was 8.4% word error averaged over 12 speakers, using the standard word-pair grammar. The system used a total of 600 sentences from a single prototype speaker and and a training sample of 40 sentences from each of the 12 test speakers. The performance of the system is comparable to several speaker-independent systems trained on 4200 sentences from 105 speakers, and tested on the same data. This result suggests that speaker adaptation may be the most cost-effective solution for applications which must be brought up quickly and must accommodate changing task domains or test conditions.</Paragraph>
  </Section>
class="xml-element"></Paper>