<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1063">
  <Title>HIGH-ACCURACY LARGE-VOCABULARY SPEECH RECOGNITION USING MIXTURE TYING AND CONSISTENCY MODELING</Title>
  <Section position="7" start_page="316" end_page="316" type="concl">
    <SectionTitle>
5. CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> New acoustic modeling techniques significantly decrease the error rate in large-vocabulary continuous speech recognition.</Paragraph>
    <Paragraph position="1"> The genonic HMMs balance the trade-off between resolution and trainability, and achieve the degree of tying that is best suited to the available training data and computational resources. For example, one can decrease the computational load by decreasing the number of genones (i.e., increasing the degree of tying) with a small penalty in recognition performance \[15\]. Our results on the various test sets represent state-of-the-art recognition performance on the 20,000-word open-vocabulary WSI task.</Paragraph>
  </Section>
class="xml-element"></Paper>