Tesprasit et al.,
C. Wutiwiwatchai, S. Furui / Speech Communication 49 (2007) 8–27 15
Author's personal copy
2003b) were based on the same idea, but used machine
learning methods to predict pauses given potential features
extracted from an input sentence. Tesprasit et al. (2003b)
applied collocation of words and the number of syllables
from the previous phrase break to the learning machine.
Both C4.5 and RIPPER have been shown to outperform
the simple POS n-gram model. Hansakunbuntheung et al.
(2005b) extended the experiment by adding other potential
features including POS context and the number of syllables
and words from previous phrase and sentence breaks.
Other types of machine learning including a neural network
(NN) and CART were also compared. The best
results were an 80% break correction rate and a 2.4%
false-break rate given by the CART engine.