<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-2003">
  <Title>Searching the Web by Voice</Title>
  <Section position="11" start_page="0" end_page="0" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have shown that a commercial speech recognition engine, using a unigram language model over words and collocations, can return the correct transcription of a spoken search query among its top 10 hypotheses about 60% of the time. Because we were not able to use a bigram model without sacrificing real-time performance, including collocations in the language model was crucial for attaining this level of recall.</Paragraph>
    <Paragraph position="1"> Still, there is a lot of room for improvement in the recall rate. One idea is to rescore the recognizer's top hypotheses with a bigram or trigram language model in a postprocessing step. However, there are many cases where the correct transcription is not among the recognizer's top 100 hypotheses. Another approach would be to adapt the acoustic and language models to individual users, but such personalization would require a different system architecture. We might also improve our language models by training on documents as well as queries (Fujii, 2001).</Paragraph>
    <Paragraph position="2"> The language models described in this paper were trained from typed queries, but queries made by voice in different settings might have quite different characteristics. For example, our data consisted of keyword queries, but voice search users might prefer to ask questions or make other types of natural language queries (which would actually be easier to model and recognize). The voice search system is currently available at labs.google.com; the data from this demonstration system could lead to improved language models in the future.</Paragraph>
  </Section>
class="xml-element"></Paper>