<?xml version="1.0" standalone="yes"?>
<Paper uid="H90-1067">
  <Title>Experiments with Tree-Structured MMI Encoders on the RM Task</Title>
  <Section position="6" start_page="349" end_page="350" type="concl">
    <SectionTitle>
CONCLUSION
</SectionTitle>
    <Paragraph position="0"> We compared vector quantization (the MD encoder) with no segmentation to a multi-stage decision-tree encoder (the MMI encoder) with and without segmentation. We found that the MMI encoder (1) extracts a significantly larger amount of information than the MD encoder; (2) works better with a  combined feature set (as a single tree); and (3) yields higher accuracy with faster decoding time when segment codes are used.</Paragraph>
    <Paragraph position="1"> In order to make a controlled comparison, neither the best decision tree technology nor the best Markov model technology was used. In decision trees, we did not use wider context in the frame tree, as in previous work \[1\]. In addition, we have found that a third segmentation stage helps, creating even larger yet &amp;quot;clean&amp;quot; segments (unpublished work at SSI). The decision tree can easily use more features simultaneously, providing the prospect of more informative codes. Since the trees make dichotomous decisions, more extensive smoothing of the codes (utilizing tree topology) should help. Further, several iterations of the entire process of labelling the frames and tree-growing can be repeated to improve accuracy (as long as the resulting recognizer provides more accurate decoding than that of the previous iteration). Finally, due to temporal compression of frames and resulting data reduction, a reduced topology of the phonetic HMMs (e.g., fewer states/transitions) may yield a better fit to the segment codes. Future research will include trying some of these variations.</Paragraph>
    <Paragraph position="2"> In our experiments, we have not fully explored the context dependency of the phonetic models. Further investigation of the use of MMI encoders with context-dependent HMMs will be conducted in the future.</Paragraph>
  </Section>
class="xml-element"></Paper>