Maybe it is the problem of adaptation speaker pattern in Config.pm. You can find the cue in the log of the estimation of mjb.SAT+dec_feat5.
Hi,
I ran several times the demo script with clb, slt, and my own data this way: I used CMU data and around 700 sentences from my own data for the training data, and 100 sentences of these 700 sentences for target speaker (I tried this way because I do not have the same sentences as the CMU ones). I did not change the settings except for the training and adaptation data. Every time I run the script I get the following error:
=====================================================================================
Start converting mmfs to the hts_engine file format at Mon May 28 11:44:30 IST 2012
=====================================================================================
/usr/local/HTS-2.2beta/bin/HHEd -A -B -C /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/configs/cnv.cnf -D -T 1 -p -i -H /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/models/qst002/ver1/cmp/mapmmf/re_clustered_sat_all.mmf -H /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/models/qst002/ver1/cmp/regTrees/dec.base -H /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/models/qst002/ver1/cmp/regTrees/dec.tree /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/edfiles/qst002/ver1/cmp/cnv_mgc.hed /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/models/qst002/ver1/cmp/tiedlist
HTK Configuration Parameters[2]
Module/Tool Parameter Value
# NATURALWRITEORDER FALSE
# NATURALREADORDER TRUE
HHEd
60375/44726 Models Loaded [7 states max, 2 mixes max]
TR 2
Adjusting trace level
// load adaptation transforms
AX /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/models/qst002/ver1/cmp/xforms/mjb.SAT+dec_feat5
Setting HMMSet adapt XForm
ERROR [+7035] Failed to find macroname mjb.SAT+dec_feat5
FATAL ERROR - Terminating program /usr/local/HTS-2.2beta/bin/HHEd
Error in /usr/local/HTS-2.2beta/bin/HHEd -A -B -C /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/configs/cnv.cnf -D -T 1 -p -i -H /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/models/qst002/ver1/cmp/mapmmf/re_clustered_sat_all.mmf -H /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/models/qst002/ver1/cmp/regTrees/dec.base -H /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/models/qst002/ver1/cmp/regTrees/dec.tree /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/edfiles/qst002/ver1/cmp/cnv_mgc.hed /home/fred/peh/HTS-demo_CMU-ARCTIC-ADAPT/models/qst002/ver1/cmp/tiedlist
Moreover, the generated waveforms are really bad. Seeing that they seemed to have too high values, I tried to decrease the volume of my waves before using them, and I still got bad results, and unusable sounds. Why do I miss mjb.SAT+dec_feat5 macroname? How should I do the adaptation knowing that my data are not the same sentences as CMU data?
Regards,
Pierre-Edouard