Sphinx4 very less accuracy

[Anurag Jain]

Hi all,

I am doing speech recognition with sphinx4. I created language model with the
help of cmuclmtk(version 0.7) toolkit and trained acoustic model using
sphinxtrain(version 1.0.7).
After training, I decoded some test files and it gave arounf 80% accuracy. But
when I did speech recognition using sphinx4 recogniser on the same files, it
gave around 10% accuracy.
I tried changing config.xml but it didn't improved accuracy much.
Please help me in this.

Regards
Anurag

[Pranav Jawale]

Hi,

After training, I decoded some test files and it gave arounf 80% accuracy.

What gave 80 % accuracy? What are you comparing against?

[Anurag Jain]

Hi,

I decoded using decode script in the scripts folder.

However I think that the problem is that I have very less amount of training
data because of which accuracy is very less.
Now I am doing acoustic model adaptation to improve accuracy.

Thanks for your reply

Regards
Anurag

[Nickolay V. Shmyrev]

If sphinx4 doesn't recognize audio while database test shows good acccuracy it
means your feature extraction doesn't match the frontend configuration in
sphinx4. Most likely your source files for acoustic model training had
incorrect format (wrong sample rate) but there could be other issues.