This is a narrowband (8kHz) Sphinx model using downsampled 
wsj_all training set.  It uses 8000 tied-states, 256
codebook Gaussians, 5 states with skip transitions and
40 CMU phones.

I downsampled the original audio using sox:
sox <original sph> -r 8000 <target> resample -ql

Feature vectors changed the following parameters to make_feat.pl
(you'll want to use the same values to the decoder):
  nfft      256
  nfilt     31
  lowerf    200
  upperf    3500
  srate     8000

Results comparing narrow and wideband using prior CMN, 
in << xRT (on a 3GHz desktop using a 10K-vocab 2-gram 
language model trained on LM-CSR):

Test set     WER 8kHz  xRT 8kHz  WER 16kHz  xRT 16kHz
--------     --------  --------  ---------  ---------
Nov92        17.63     0.0910    15.38      0.0839
Hub1         31.84     0.1141    29.08      0.1012
Hard         36.48     0.1286    33.38      0.1011

Nov92 = WSJ0 SI_ET_05, 330 utts
Hub1  = WSJ0 SI_DT_20, WSJ1 SI_DT_20, 894 utts
Hard  = WSJ SI_DT_JR, WSJ1 SI_DT_S2/SJM SI_ET_S2/SJM SI_ET_H!/WSJ64K, 784 utts

Have fun!
Keith Vertanen