| Summary: | 碩士 === 國立清華大學 === 統計學研究所 === 86 === In this thesis, we use HHM to deal with the problem of large
vocabulary recognition task of Taiwanese. The phone units we
use are right context dependent (RCD) phonemes. If we just
consider the effect of inside syllable context dependency, the
top 1 recognition accuracy reaches 88%. If we modify our models
to include all inter-syllable context dependency, the result of
top 1 recognition accuracy increases to 92.11% but the total
number of states in our network increases to 4 times. To
compensate the insufficient training data due to the large
amount of models, we use the method of Acoustic Decision Tree to
do state clustering.The total number of states decreases to 3/5,
and the recognition rate also increases to 92.87%.
|
|---|
