[hts-users:02418] What is the usage of a cascade of CMLLR transforms?

[Hui LIANG <tshlmail-hts@xxxxxxxxx>]

Hello everyone,

I was doing speaker adaptation using CMLLR and trying to involve a cascade of transforms in my experiments. Specifically, I had a set of fixed parent transforms and wanted to update input transforms only. According to my understanding of the HTK Book, I tried to utilize parent transforms in the following way:

The file of my parent transforms:

~a "cmp.cmllrParent"
<ADAPTKIND>TREE
<BASECLASS>~b "dectree_cmp_parent.base"
<XFORMSET>
<XFORMKIND>CMLLR
<NUMXFORMS> 125
<LINXFORM> 1
....

The file of my input transforms:

~a "spkr.cmp.cmllr0"
<ADAPTKIND>TREE
<BASECLASS>~b "dectree_cmp.base"
<PARENTXFORM> ~a "cmp.cmllrParent"
<XFORMSET>
<XFORMKIND>CMLLR
<NUMXFORMS> 625
....

However, in the resulting output transform file of the first iteration, the line with the macro "<PARENTXFORM>" disappeared. I added the line to the output transform file manually so the file finally looked like:

~a "spkr.cmp.cmllr1"
<ADAPTKIND>TREE
<BASECLASS>~b "dectree_cmp.base"
<PARENTXFORM> ~a "cmp.cmllrParent"    # I added this line manually, which was absent from my output transform file.
<XFORMSET>
<XFORMKIND>CMLLR
<NUMXFORMS> 619
....

Then this modified file was given to HERest as an input transform file for the second iteration.

What I didn't understand is that the likelihoods of adaptation utterances in the second iteration were significantly lower than those in the first iteration. So I suspect whether I was utilizing parent transforms in a correct way. (I had also tried the "-E" and "-W" switches of HERest but they seemed to work only for updating model parameters during speaker adaptive training.)

My understanding is that parent transforms are applied to features (because of CMLLR) first and that input transforms are then applied to the adjusted features. Hope the understanding is correct.

Thanks a lot,
Hui LIANG