<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-1028">
  <Title>Towards Automatic Scoring of Non-Native Spontaneous Speech</Title>
  <Section position="12" start_page="1" end_page="1" type="concl">
    <SectionTitle>
10 Conclusions and future work
</SectionTitle>
    <Paragraph position="0"> This paper is concerned with explorations into scoring spoken language test items of non-native speakers of English. We demonstrated that an extended feature set comprising features related to length, lexical sophistication, fluency, rate and content could be used to predict human scores in SVM models and to illuminate their distribution into five different classes by means of a CART analysis.</Paragraph>
    <Paragraph position="1"> An important step for future work will be to train the acoustic and language models of the speech recognizer directly from our corpus; we are additionally planning to use automatic speaker adaptation and to evaluate its benefits. Furthermore we are aware that, maybe with the exception of the classes related to fluency, rate and length, our feature set is as of yet quite rudimentary and will need significant expansion in order to obtain a broader coverage of communicative competence.</Paragraph>
    <Paragraph position="2"> In summary, future work will focus on improving speech recognition, and on significantly extending the feature sets in different categories.</Paragraph>
    <Paragraph position="3"> The eventual goal is to have a well-balanced multi-component scoring system which can both rate non-native speech as closely as possible according to communicative criteria, as well as provide useful feedback for the language learner.</Paragraph>
  </Section>
class="xml-element"></Paper>