<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1062">
  <Title>Tree-Based State Tying for High Accuracy Modelling</Title>
  <Section position="6" start_page="310" end_page="310" type="concl">
    <SectionTitle>
5. CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> This paper has described an efficient method of state clustering based on the use of phonetic decision trees and its use has been demonstrated in the HTK tied-state recognition system. It has been shown that tying at the state rather than the model level gives improved accuracy and that phonetic decision trees are as effective for clustering as data-driven methods but have the key advantage of providing a mapping for unseen triphones.</Paragraph>
    <Paragraph position="1"> The overall results on both the RM and WSJ tasks indicate that the proposed approach leads to a recogniser with state-of-the-art performance but which is relatively compact and easy to construct. The method depends crucially on the use of continuous density HMMs since they provide a simple way of manipulating complexity.</Paragraph>
    <Paragraph position="2"> Initially when the data for some triphones is sparse, the use of simple single Gaussian distributions still allows reasonable parameter estimates to be made. The use of single Gaussians in the initial stages also allows very efficient tree-building since the required likelihood-based objective function can be computed without reference to  20k task using bigram (bg) and trigram (tg) language models, t denotes systems used for the ARPA November 1993 WSJ evaluation.</Paragraph>
    <Paragraph position="3"> the training data. However, once the amount of data per state has been increased by the state tying procedure, the single Gaussians can easily be converted to mixture Gaussians by splitting components and re-estimating.</Paragraph>
    <Paragraph position="4"> Model complexity can then be increased smoothly in this way until optimal performance is achieved.</Paragraph>
  </Section>
class="xml-element"></Paper>