<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-2002">
  <Title>Factored Language Models and Generalized Parallel Backoff</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Discussion
</SectionTitle>
    <Paragraph position="0"> The improved perplexity bigram results mentioned above should ideally be part of a first-pass recognition step of a multi-pass speech recognition system. With a bigram, the decoder search space is not large, so any appreciable LM perplexity reductions should yield comparable word error reductions for a fixed set of acoustic scores in a firstpass. For N-best or lattice generation, the oracle error should similarly improve. The use of an FLM with GPB in such a first pass, however, requires a decoder that supports such language models. Therefore, FLMs with GPB will be incorporated into GMTK (Bilmes, 2002), a general purpose graphical model toolkit for speech recognition and language processing. The authors thank Dimitra Vergyri, Andreas Stolcke, and Pat Schone for useful discussions during the JHU'02 workshop.</Paragraph>
  </Section>
class="xml-element"></Paper>