Low Accuracy

[Ramshad]

Hi,
I would like to make a blog post to thank the Admins for the fabulous blog.

Now i am using Openears Library Version0.913 – which is released on October
11th, 2011 for Speech recognition.

My clients are from California.So they told me a concern about the accuracy.

for improving it, i have found the way to Adapting the default acoustic model
from the below site

null

I have done all the steps successfully and created acoustic model.

I have trained the acoustic model with arcticAll first 593 sentences with my
own voice.

I didn't get any warnings and errors when created the acoustic model.

I have used standard ARPA format language model file for Openers.
my .dic file contains only 40 words..dic file contends are

ABOVE AH B AH V
ADD AE D
BEGINNING B IH G IH N IH NG
BETWEEN B IH T W IY N
BETWEEN(2) B IY T W IY N
CIRCLE S ER K AH L
CLEAN K L IY N
DEACTIVATE D IY AE K T IH V EY T
DIRTY D ER T IY
DOWN D AW N
EMPTY EH M P T IY
EMPTY(2) EH M T IY
EXERCISE EH K S ER S AY Z
FALL F AO L
FALL(2) F AA L
FINISHED F IH N IH SH T
FRONT F R AH N T
GRAB G R AE B
GREEN G R IY N
HEAVY HH EH V IY
HELP HH EH L P
IN IH N
INHALE IH N HH EY L
KICK K IH K
KISS K IH S
LAST L AE S T
LAST(2) L AO S T
LAST(3) L AE S
LEAVE L IY V
MAD M AE D
MIDDLE M IH D AH L
NARROW N EH R OW
NARROW(2) N AE R OW
NIGHT N AY T
OF AH V
ORANGE AO R AH N JH
ORANGE(2) AO R IH N JH
OUTSIDE AW T S AY D
PICK P IH K
PLAY P L EY
POINT P OY N T
RAINY R EY N IY
RECTANGLE R EH K T AE NG G AH L
SAD S AE D
SHARP SH AA R P
SKATING S K EY T IH NG
TICKLE T IH K AH L
UNZIP AH N Z IH P
UP AH P
VACUUM V AE K Y UW M
WRITE R AY T

However i found only 60-70 % accuracy with this acoustic model.

i have discussed this topic earlier in Openears blog

http://www.politepix.com/forums/topic/clarificatios-need-about-
openears/#post-8382

From the above .dic file, CIRCLE,TICKLE,LEAVE words are not detected when i
attempted 10 times.

POINT always detected as ORANGE!!!

ADD always detected as MAD!!

and other words detected 6-7 times when attempted 10 times.

I have used PocketSphinx 0.7 SphinxBase 0.7, SphinxTrain 1.0.7 for
training.

would any one please suggest me better tips for improving the accuracy

is 593 arctic sentences enough for training??

Thanks

[Nickolay V. Shmyrev]

Hello

The accuracy you are describing doesn't seem good. Most likely you have done
some mistake in adaptation. Please provide the adaptation data you were using
and the adaptation scripts. You can upload them to a public file sharing
resource.