Create M/F models (used for VTLN)

First, create a male/female segmentation of the data using the information on the speakers provided in the resources (we use only WSJ0 for this):

cd exp
../scripts/make_mf_seg.py ../resources/SPRAAK/wsj_si84_train.cor "../resources/SPRAAK/wsj0-spkr-info.txt.920128 ../resources/SPRAAK/wsj0-spkr-info.txt.add"  ../data/wsj0/ wv1 > ../resources/wsj_si84_train_vtln.seg
spr_train.py wsj_vtln_128.config

The models trained here are quite simple, including noise in both male and female models. This might affect M/F modeling (and consequently VTLN processing) in files that have long silence parts; however, in the case of the WSJ database this will have little effect.

HMM training

The configuration files for training acoustic models can be found in the ./scripts directory.

The configuration file wsj0+1_mida_vtln_CD.config will create high quality context dependent models.

The training is started by typing:

spr_train.py wsj0+1_mida_vtln_CD.config >& wsj0+1_mida_vtln_CD.ERR