Programmed by Chanwoo Kim
 for ICASSP 2010
Oct. 12, 2009
modified Feb. 11, 2011

1) Just Run PNCC_ICASSP2010('outFile', 'inPutFile')



IMPORTANT : The input is assumed to be single-channel NIST Sphere format OR
Microsoft WAV. The input file extension is automatically read by the program, but the header is not checked, it's just skipped. 

The sampling rate is assumed to be 16kHz.

We used this program in getting result in ICASSP 2010paper on the RM1 and WSJ0 database



2) Output is a feature in Sphinx format





3) To see the spectrogram demo

  Type DemoBatch.m in Matlab

  It will launch three figures: 
  a) spectrogram, 
  b) spectrogram using log nonliearity (This one is still better than MFCC in our recognition experiment)
  c) gammatone weighting, PNCC (This one is much better than b))

  In each figure you will sub-figures obtained under different noisy conditions. If the spectrogram from the noisy
  condition is closer to the top subfigure in each figure, then we consider that feature to be more robust