help for increasing the accuracy for a content based audio classification syste

Hi;
We are doing the above mentioned project for audio classification.We
are coding in MATLAB.We have tried implementing the code by increasing
the number of classes in the database from 4 to 7 and also the number
of MFCC's from 13 to 30.We stuck to using left right HMM for now.In
spite of all this the system accuracy remains more or less
unaffected.By accuracy I mean the percentage of clips that are getting
correctly classified.Just to make it more clear here is the gist of
the result:

class accuracy
1) Male speech: 90%
2) female speech; 88%
3) music: 72%
4) animal sounds: 48%
5) nature sounds : 58%
6) violence: 52%
7) traffic sounds: 38%

The database has been collected over the net as it is impossible
to record animal and nature sounds.We are trying to append ZCR to the
existing list of features but we are not sure if it will suffice. Is
there any other way to improve this problem. Will extracting features
like centeroid and spectral flux help?If there is something that may
work out please do let us know. Start a New Thread