I'm trying to do some discriminative training experiments with both
CDHMMs and SCHMMs. For this I use HVite to generate the lattices for
HMMIRest and it works fine for the continuous density models. For the
semi-continuous models however, HMMIRest breaks during the update
process.

My process is as follows:

I did some simple ML models and after training them to continuous
monophone models I used HHEd to convert them to a tied-mixture set
using this HHEd edit script (which was suggested in both the HTK Book
and the samples):

JO 256 2.0
TI TM {*.state[2-4].mix}
HK TIEDHS

The resulting models work fine when using them with HERest for normal
ML re-estimation. However, with HMMIRest these models don't work. I
call HMMIRest using this call:

HMMIRest -H hmms/hmm20/macros -H hmms/hmm20/hmmdefs -S scpfile -q
dt/plat.num -r dt/plat.den -u tmvw -M hmms/hmm21 modelfile

The error it throws is:

Program received signal EXC_BAD_ACCESS, Could not access memory.
Reason: KERN_PROTECTION_FAILURE at address: 0x00000004
Do MixUpdate (mp=0x1bf978, s=1, Lr=0.000104866849, meescale=1, t=1) at
HFBLat.c:1041

My worries are that HMMIRest will not function with TIEDHS marked HMM
sets. If there is no way to fix this problem right now, can I at least
go around it somehow? For example by not marking the HMM set as a
"TIEDHS" system, but tying the mixtures nonetheless. As I understand
it, using the "HK TIEDHS" command is just for reducing the memory
footprint of the HMM set.