Although Thai ASR has been widely investigated, there
is still a lack of research at state-of-the-art level; likewise,
there has been no benchmark evaluation of LVCSR. One
of the most important obstacles delaying the development
of a Thai LVCSR is the lack of a large annotated text
corpus for language modeling. Constructing such a corpus
requires an accurate word segmentation tool, which could
shorten the hand-correction process. Current segmentation
tools produce more than 90% accuracy but the accuracy is
significantly lower when applied to text in new domains
(Aroonmanakun, 2002). Due to the difficulty currently
encountered in constructing LVCSR engines, researchers
have tended to focus on other Thai-specific issues, such
as tones. Tone analysis and modeling itself is useful for
TTS, but provides little improvement for ASR as shown
in many publications on Chinese. The current state-ofthe-art
TTS, based on variable-length unit-selection, has
been carried out at only one research site, NECTEC.
Hence, the advancement of Thai TTS remains slow. Considering
the aforementioned limitations, the following are
significant issues the authors would suggest Thai researchers
to turn their attention to: