<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3005">
  <Title>Automatic Call Routing with Multiple Language Models</Title>
  <Section position="6" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
5. Discussion
</SectionTitle>
    <Paragraph position="0"> In this paper, we have presented a method for automatic call routing in which we do not require transcriptions of the training utterances, only the route of each utterance.</Paragraph>
    <Paragraph position="1"> The technique is based on phonetic recognition of utterances, and we have focused on the design of the language model in this recognition process. Our conclusions are that iterating a single phone language model (as described in [1]) is highly beneficial to performance, but performance can be further increased by using multiple language models for recognition for utterances whose content is ambiguous when a single language model is used. Using multiple LMs inevitably gives rise to identification of false keywords, but this difficulty is resolved by the use of post-processing HMMs which estimate the likelihood of the putative keyword phonetic sequence being present in the waveform. Future work will concentrate on use of confidence measures and classification of ambiguous utterances. We will also investigate the use of &amp;quot;lightly supervised&amp;quot; adaptation, in which a small proportion of the utterances available have been transcribed [15].</Paragraph>
  </Section>
class="xml-element"></Paper>