Making CMU Sphinx recognize a simple fake language

[Antidamage]

I'm building a VR application that relies on voice input to cast spells. My fictional spell-casting language has 30 short words (each is a single syllable) that combine to make a spell phrase. Sometimes I get good recognition results from validation using a grammar file but it's still fairly inaccurate a lot of the time.

What do I do now? Should I create a new language model for my (extremely small) language? I understand that creating a larger language model takes a long time, but this one is really dead simple and short.

Or can an accousting model adaptation help me?

To give you an idea of what I'm working with here's my entire spell language (although I intend to change some of the similar-sounding terms):

tor T AO R
don D AO N
vos V AO S
kai K AY
vad V AA D
zor Z AO R
sko S K OW
hut HH AH T 
vash V AE SH
uth UW TH
huk HH AH K
bur B ER
ket K EH T
han HH AE N
soth S AO TH
ara AA R AH
go G OW
ro R OW
na N AA
ke K EH
haya HH AA IY AA
mo M OW
se S EH
sha SH AA
ge G EH
toh T AO HH
var V AA R
en EH N
de D EY
cho CH OW

[Nickolay V. Shmyrev]

To get help on the accuracy you need to provide all the data to reproduce your accuracy issues - audio, data, reference transcript. You can check

http://cmusphinx.github.io/wiki/tutorialtuning

for details