For the TANDEM system, we could either train a set of HMM models on the projected posterior features, we could also train them directly on the posteriors and then using HTK to do a HLDA projection to reduce the dimension of the feature.
However, there is a limitation to the dimension of the feature vectors used: the dimension should be smaller than 100.
This is due to the HMath.c file, in which, there are several functions have defined some local array with the fixed length 100.
For my copy of the v3.4.1 HTK, those are:
1424: float col;
1427: int n,i,j,perm;
1477: double col;
1480: int n,i,j,perm;
1506: double col;
1509: int n,i,perm;
1535: double col;
1538: int n,i,perm;
What needs to do is just increase the size of those local arrays.