[hts-users:02905] Problem with speaker adaptation using HTK
Hi!
I am trying to adapt a set of trained HMMs to a particular speaker. The characteristics of the system are:
1. The HMMs are trained at a phone level on English sentences only.
2. The number of monophones are 46.
3. Each phone HMM is modelled with 5 states (3 emitting states). Each state in turn is modelled with 25 mixtures.
4. Each of the mixtures in the above statement is in turn modelled with a 39 dimensional Gaussian. The input wave is converted to MFCC_E_D_A parameters.
The above system is being adapted to a particular speaker. Please note that I am trying to adapt the monophone level HMMs.
Adaptation process proceeds like so:
1. Took about 200 sentences of the desired speaker and converted it into mfcc using HCopy.
2. Ran HVite to perform a forced alignment of the data. HVite was invoked like so:
HVite -A -D -T 1 -l '*' -o SWT -b SHVite -A -D -T 1 -l '*' -o SWT -b SENT-END -C ../config.txt -H hmm_englishOnly_4_orig/hmmdefs -H hmm_englishOnly_4_orig/macros -i adaptation.mlf -m -y lab -a -I trainingData/train1/sandeep_Words.mlf -S scripts/train1_mfccList.scp dict_harshat monophones1eng
3. Ran HHEd to build a regression class tree and get a transformation. HHEd invoked like so:
HHEd -A -D -T 1 -H hmm_englishOnly_4_orig/macros -M classes regtree.hed
monophones1engH hmm_englishOnly_4_orig/hmmdefs -H
hmm_englishOnly_4_orig/macros -M classes regtree.hed monophones1eng
where: hmm_englishOnly_4_orig/ contains the monophone hmms.
regtree.hed:
RN "models"
LS "stats_engOnly_3_4"
RC 32 "rtree"
"stats_engOnly_3_4" are state occupation stats file.
On running HHEd in the above fashion, the error I get is as follows:
Splitting Node 34, score 7.707355e+04
Splitting Node 18, score 7.067691e+04
Splitting Node 25, score 6.851620e+04
Splitting Node 32, score 6.397755e+04
ERROR [+999] Components missing from Base Class list (2413 3375)
ERROR [+999] BaseClass check failed
FATAL ERROR - Terminating program HHEd
I have not been able to decode the above errors. I have tried the following so far:
1. I am trying to step through the source code for HTK and see if there is a problem somewhere. Have had little success so far.
2. I notice in the hmm definition files for the monophones that a lot of the mixtures are missing. I am assuming that these mixtures have zero weight and hence have been omitted in the definition file. Could this error be because of these components 'missing'? I tried to insert a mixture wherever it was missing with zero weights. That did not help. The above error still persists.
Any help in making this go forward is most welcome.
Regards,
Sriram Shankar.