My accuracy for GSM and CELP(4.8kbs) Is Low. Why?

im new to speech coding. Currently im trying to calculate speech
recognition accuracy for TIMIT corpus which is coded using GSM and
CELP 4.8kbps coder. im using HTK toolkit.

i find out the my speech recognition accuracy drop drastically when
i tested coded speech using the acoustic model trained using NIST
(wav) file from TIMIT corpus.below i list down my method and can
someone of you please clarify whether i done something wrong ? im
doing phoneme recognition system.

1. i read Timit(nist) using matlab

2. resample it to 8khz 16 bit

3. then enter the resampled file as an input to codectest.c
(HAWKCODER)

4. i get GSM and CELP 4.8 kb/s (lpc) as output in the format
of '.au'.

5. then convert the GSM and CELP (au file) to MFCC_0_D_A (39) using
Hcopy

from HTK (using config file)

6. later i insert the feature vectors to the trained acoustic model.
i get results
WORD: %Corr!.22, Acc.36 for GSM

but when i tested same TIMIT (Nist WAV) withoud coding

WORD: %Corrr.76, AccX.43

I just wan to know whether i had done the steps correctly or i had
done something wrong? why my accuracy drop drastically? do you have
any matlab code for GSM and CELP coder? or any info where i can it?

Waiting for your reply.Thank you for your help.

Regards

Nash Start a New Thread