Accuracy unchanged after training. After adaption even worse

[Sudomatze]

Hi there,

like the titel said, i have a problem with the accuracy.

I use the newest german voxforge language model which is already trained with lots of data and I want to train it with some own audios. At the moment I only have 100 audios, each whith one sentance and yes I know it's a way too small database.

So I read that adaption is a better choice if the database is very small so I tried. The accuracy before doing anything was at 58.03% and after adaption it was about 7%

Then I tried the sphinxtrain (maybe this would work better) but in this case the accuracy was completely unchanged. Is the reason the small amount of data?

Tell me if you need some additional informatino to help me.

thanks for helping.

edit:
after the training there is the new folder XXX.ci_cont in the model_parameters folder. This should be the trained model but if I test this one the accuracy is 18,65 %. So it is like the adaption worse than before

[Sudomatze]

oh and the way I doing the test is with this code:

pocketsphinx_batch \
 -adcin yes \
 -cepdir wav \
 -cepext .wav \
 -ctl etc/XXX_test.fileids \
 -lm etc/XXX.lm.bin \
 -dict etc/XXX.dic \
 -hmm model_parameters \
 -hyp test.hyp;
perl ./word_align.pl etc/XXX_test.transcription test.hyp > test.txt;

and my sphinx_train.cfg is attached

[Sudomatze]

Does anybody know something about my problem?
I'm not sure if I'm doing something wrong or just have a too small amount of data.