|
This research is focused on resolving the speech recognition
pro blem for highly confusing abrupt-tone syllables of
Taiwanese Hokkian. A series of analytic experiments are
implemented to investigate certain characteristics of abrupt-
tone syllables in Taiwanese Hokkian. From the spectral
analysis, it can be conclu- ded that the ending part of abrupt-
tone syllables represents not an individual phoneme,but an
shifting version of non-abrupt-tone one. There are totally 37
syllables containing abrupt-tone and non-abrupt-tone sets being
selected for the recognition experi- ments. Different features
including cepstrum, mel-cepstrum, and LPC code with HMM models
for these confusing syllables from a speaker-dependent to a
speaker-independent mode are used to test the recognition
performance. Also, a new feature called modified mel-cepstrum
with emphasis in the band of 1K to 4K Hz is applied to the
experiments. At best,the speaker-independent recognition rate
for the whole database is over70%. the top two recognition
rates can improve to 87%. For abrupt-tone syllables only, the
recognition rate is about 65% for top one, and 82.8% for top
two. The new feature of modified mel-cepstrum performs well in
a speaker-independent mode with 81.6% top two recognition rate
in comparison with 79.2% of mel-cepstrum. In a summary,
cepstrum outperforms the other three features for the
recognition of the abrupt-tone syllables in Taiwanese Hokkian,
but the difference with mel-cepstrum or modified mel-cepstrum
is less than 3%.
|