<?xml version="1.0" standalone="yes"?>
<Paper uid="P91-1039">
  <Title>FACTORIZATION OF LANGUAGE CONSTRAINTS IN SPEECH RECOGNITION</Title>
  <Section position="8" start_page="303" end_page="304" type="evalu">
    <SectionTitle>
4. Experimental Results
</SectionTitle>
    <Paragraph position="0"> The semantic postproeessor was tested using the speech recognizer arranged in different accuracy conditions. Results are summarized in Figures 1 and 2. Different word accuracies were simulated by using various phonetic unit models and the two covering grammars (i.e. NG and WP). The experiments were performed on a set of 300 test sentences known as the February 89 test set (Pallett. 1989) The word accuracy, defined as 1- insertions deletions'e substitutions xl00 (3) number of words uttered was computed using a standard program that provides an alignment of the recognized sentence with a reference string of words. Fig. 1 shows the word accuracy after the semantic postprocessing versus the original word accuracy of the recognizer using the word pair grammar. With the worst recognizer, that gives a word accuracy of 61.3%, the effect of the semantic postprocessing is to increase the word accuracy to 70.4%. The best recognizer gives a word accuracy of 94.9% and, after the postprocessing, the corrected strings show a word accuracy of 97.7%, corresponding to a 55% reduction in the word error rate. Fig. 2 reports the semantic accuracy versus the original sentence accuracy of the various recognizers. Sentence accuracy is computed as the percent of correct sentences, namely the percent of sentences for which the recognized sequence of words corresponds the uttered sequence. Semantic accuracy is the percent of sentences for which both the sentence generation template and the values of the semantic variables are correctly decoded, after the semantic postprocessing. With the best recognizer the sentence accuracy is 70.7% while the semantic accuracy is 94.7%.</Paragraph>
    <Paragraph position="1">  cessing When using acoustic verification instead of simple phonetic verification, as described in  section 3.2, better word and sentence accuracy can be obtained with the same test data. Using a NG covering grammar, the final word accuracy is 97.7% and the sentence accuracy is 91.0% (instead of 92.3% and 67.0%, obtained using phonetic verification). With a WP covering grammar the word accuracy is 98.6% and the sentence accuracy is 92% (instead of 97.7% and 86.3% with phonetic verification). The small difference in the accuracy between the NG and the WP case shows the rebusmess introduced into the system by the semantic postprocessing, especially when acoustic verification is peformed.</Paragraph>
  </Section>
class="xml-element"></Paper>