Recognition using any one of the bands

[karthik Pandia]

I wanted to see the recognition accuracy on individual bands. 

To do this, I replaced merger norms and weights with the band0/1 norms and weights.

When I use band0, irrespective of the utterance I give as input, the full file is recognized as  /k/.

Similarly /t/ for band1. Moreover, the log likelihoods I get is far lesser than I get using the merger net.

I am wondering if this is the right way to use the individual nets? 

-Karthik