<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2907">
  <Title>tioning: Low latency real-time broadcast news tran-</Title>
  <Section position="8" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
6.3 Results
</SectionTitle>
    <Paragraph position="0"> We implemented some of the proposed techniques and made comparisons with the previous method used by Saraclar and Sproat (2004). The full indexing method consumed too much time while indexing Broadcast News lattices and used too much memory while indexing phone lattices for Teleconferences. In the other cases, we confirmed that the new method yields identical results. In Table 1 we compare the index sizes for full indexing and partial indexing with the previous method. In both cases, the input lattices are pruned so that the cost (negative log probability) difference between two paths is less than six.</Paragraph>
    <Paragraph position="1"> Although the new method results in much smaller index sizes for the string case (i.e. nbest=1), it can result in very large index sizes for full indexing of lattices (cost=6).</Paragraph>
    <Paragraph position="2"> However, partial indexing by length restriction solves this problem. For the results reported in Table 1, the length of the word strings to be indexed was restricted to be less than or equal to four, and the length of the phone strings to be indexed was restricted to be exactly four.</Paragraph>
    <Paragraph position="3"> In Saraclar and Sproat (2004), it was shown that using word lattices yields a relative gain of 3-5% in maxF over using best word hypotheses. Furthermore, it was shown that a &amp;quot;search cascade&amp;quot; strategy for using both word and phone indices increases the relative gain over the baseline to 8-12%. In this strategy, we first search the word index for the given query, if no matches are found we search the phone index. Using the partial indices, we obtained a precision recall performance that is almost identical to the one obtained with the previous method. Comparison of the maximum F-measure for both methods is given in  corpora.</Paragraph>
    <Paragraph position="4"> As an example, we used a filter that indexes only consonants (i.e. maps the vowels to epsilon1). The resulting index was used instead of the full phone index. The size of the consonants only index was 370MB whereas the size of the full index was 431MB. In Figure 5 we present the precision recall performance of this consonant only in-</Paragraph>
  </Section>
class="xml-element"></Paper>