<?xml version="1.0" standalone="yes"?>
<Paper uid="H91-1028">
  <Title>FIELD TEST EVALUATIONS and OPTIMIZATION of SPEAKER INDEPENDENT SPEECH RECOGNITION for TELEPHONE APPLICATIONS</Title>
  <Section position="6" start_page="162" end_page="162" type="concl">
    <SectionTitle>
CONCLUSION
</SectionTitle>
    <Paragraph position="0"> Exhaustive analysis of field trials allowed to better identify the most crucial shortcomings of the speech recognition systems developed in the laboratory and to substantially improve both the speech recognition performance and the acceptability of the resulting services. From our experience, it appears that both the rejection of incorrect inputs and the noise-speech end-point detection are among the most crucial problems.</Paragraph>
    <Paragraph position="1"> A new reiection procedure has been presented which still requires turther refinements. Introducing field data in the training database proves to be an efficient procedure for rapidly improving the performances of systems that can be re-trained during their exploitation. Recognition score improvements were obtained by increasing the number of acoustic coefficients and HMM model parameters.</Paragraph>
    <Paragraph position="2"> Finally, dynamic adjustments of the structure of Markov models allowed to either reduce the overall model complexity (a crucial point for industrial implementations) or improve the recognition performance especially for larger vocabularies where the use of sub-word basic units becomes necessary.</Paragraph>
  </Section>
class="xml-element"></Paper>