<?xml version="1.0" standalone="yes"?>
<Paper uid="H93-1022">
  <Title>IMPROVED KEYWORD-SPOTTING USING SRI'S DECIPHER TM LARGE-VOCABUARLY SPEECH-RECOGNITION SYSTEM</Title>
  <Section position="7" start_page="116" end_page="117" type="ackno">
    <SectionTitle>
5. SUMMARY
</SectionTitle>
    <Paragraph position="0"> This paper describes how SRI has applied our speaker-independent large-vocabulary CSR system (DECIPHER TM) to the keyword-spotting task. A transcription is generated for the incoming spontaneous speech by using a CSR system, and any keywords that occur in the transcription are hypothesized. We show that the use of improved models of non-keyword speech with a CSR system can yield significantly improved keyword spotting performance.</Paragraph>
    <Paragraph position="1"> The algorithm for computing the score of a keyword combine information from acoustic, language, and duration. One key limitation of this approach is that keywords are only hypothesized if they are included in the Viterbi baektrace. This does not allow the system builder to operate effectively at high false alarm levels if desired. We are eousidering other algorithms for hypothesizing &amp;quot;good score&amp;quot; keywords that are on high scoring paths. We introduced an algorithm for smoothing language model probabilities. This algorithm combines small task-specific language model training data with large task-independent language training data, and provided a 14% reduction in test set perplexity. null The use of a large-vocabulary continuous-speech recognition system allows the system designer a great dealof flexibility in choosing the keywords that they would like to select for the particular application. If the desired keyword is already in the lexicon, then searching for the keyword can be achieved by looking for the word in the transcription generated by the recognizer. If the word is not in the lexicon, the word can be easily added to the system since triphone models have already been trained.</Paragraph>
    <Paragraph position="2"> The ability to transerihe spontaneous speech and search for relevant keywords will play an important role in the future development of simple spoken language applications. Such systems will be easily portable to new domains. Since the operating point for our speech recognizer is typically one which has a low insertion rate, there is little chance for a keyword false alarm.</Paragraph>
    <Paragraph position="3"> Future experimentation will determine the effectiveness of such understanding systems for human-computer interaction.</Paragraph>
  </Section>
class="xml-element"></Paper>