<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3005">
  <Title>Automatic Call Routing with Multiple Language Models</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2. Database
</SectionTitle>
    <Paragraph position="0"> The application studied here was the enquiry-point for the store card for a large retail store. Customers were invited to call up the system and to make the kind of enquiry they would normally make when talking to an operator. Their calls were routed to 61 different destinations, although some destinations were used very infrequently. 15 000 utterances were available, and a subset of 4511 utterances was used for training and 3518 for testing, in which 18 different call types were represented. Some of these call types are quite easily confused e.g. PaymentDue and PaymentDate, PaymentAddress and Changeaddress. Phoneme recognition of the input speech queries was performed using an HMM recogniser whose acoustic models had been trained on a large corpus of telephone speech and which had separate models for males and females. The average length of an utterance is 8.36 words. In addition, transcriptions of the prompts from the Wall Street Journal (WSJ) database were used to generate phoneme-level statistical language models for initial training. These models were generated using a scheme for backing off to probability estimates for shorter ngrams. null The size of the vocabulary is 1208 words. To get a feel for the difficulty of the task, the mutual information (MI) between each word and the classes was calculated. By setting a threshold on this figure, we observed that there were about 51 keywords occurring in 4328 utterances which were capable on their own of classifying a call with high accuracy (some utterances had no keywords).</Paragraph>
  </Section>
class="xml-element"></Paper>