<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-1107">
  <Title>A Method for Correcting Errors in Speech Recognition Using the Statistical Features of Character Co-occurrence</Title>
  <Section position="4" start_page="656" end_page="656" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> As described above, our proposed method has the following features: (1) Since the proposed method is designed with a arbitrary length string as a unit, it is capable of correcting errors which are hard to deal with by methods designed to treat words as units.</Paragraph>
    <Paragraph position="1"> For example, the insertion error '~&amp;quot; (&amp;quot;wo&amp;quot;) in the string '3~f.~L ~,~ Jj&amp;quot;(~ ' (&amp;quot;shiharai wo houhou'~ shown in table 2-1 cannot be corrected by a method designed to treat words as units, because of the existence of the particle' ~' (&amp;quot;wo&amp;quot;) as a correct word. However with the proposed method, it is possible to correct this kind of error by using the row of characters before and after '~' (&amp;quot;wo&amp;quot;).</Paragraph>
    <Paragraph position="2"> (2) In the proposed method of learning the trend of errors and expressions with long strings, it is possible to correct errors where it is difficult to narrow the candidates down to the correct character with the probability of the character sequence alone.</Paragraph>
    <Paragraph position="3"> When considering the candidate for &amp;quot;(&amp;quot; (&amp;quot;te&amp;quot;) in' l.,U. &amp;quot;( ~ ~. ~ (c)U.&amp;quot; (&amp;quot;shitetekimasunode '~) shown in table 2-1 to satisfy the probability of the character sequence, its candidates, '4 ~' (&amp;quot;/&amp;quot;), '}3' Co&amp;quot;), 'I~' (&amp;quot;itada'~ are arranged in order of increasing probability. It is therefore difficult to narrow the candidates into the correct character 'I~' (&amp;quot;itada&amp;quot;) by the probability of character sequence alone. But with the proposed method it is possible to correct this kind of error by using the row of the characters before and after &amp;quot;(&amp;quot; Cte&amp;quot;).</Paragraph>
    <Paragraph position="4"> (3) Both the Error-Pattem-Database and String-Database can be mechanically prepared, which reduces the effort required to prepare the databases and makes it possible to apply this method to a new recognition system in a short time.</Paragraph>
    <Paragraph position="5"> From the evaluation, it became clear that the proposed method has the following effects:  (1) It reduces over 8% of the errors.</Paragraph>
    <Paragraph position="6"> (2) It improves the understanding of the recognition results by7%.</Paragraph>
    <Paragraph position="7"> (3) It has very little influence on correct recognition results. (4) It is more applicable for a recognition result with a few  errors than one with many errors.</Paragraph>
    <Paragraph position="8"> Judging from these results and features, the use of the proposed method as a post-processor for speech recognition is likely to make a significant contribution to the performance of speech translation systems.</Paragraph>
    <Paragraph position="9"> In the future, we will try to improve the correcting accuracy by changing algorithms and will also try to improve translation performance by combining our method with Wakita's method.</Paragraph>
  </Section>
class="xml-element"></Paper>