Closed ReinholdM closed 4 years ago
Hi ! We added the three kind of features so you just have to swap the name in the cfg file to try with different features. Therefore, you just have to do the normal Kaldi run.sh, and then compute the Fmllr as explained in the tutorial !
Thanks for your reply. Can I understand that I should do the run.sh
three times for three types of features and change the name respectively in the cfg file?
in the pytorch-kaldi cfg file, you can remove the features that are not used so you don't have the error anymore if they do not exist. Then you just need to run the run.sh once, and call the command to generate the fmllr (fmllr are computed in top of traditional features).
got it! It's just what I want to ask, the simple question. Forgive me for incompletely understanding this toolkit. Thanks again!
No problem! Happy that this toolkit can help your work. Note that SpeechBrain will come out soon (Fall-winter), and will be way better and easier.
Hi! There still some problems when running run_exp.py
with cfg/TIMIT_baslines/TIMIT_liGRU_fmllr.cfg
.
I delete the mfcc
and fbank
fea_lst section and only change the path of fmllr
feats.scp, as well as specify the fea_opts as follows
But error information like this happened
So how can I fix it next? Thx
I want to follow the config in
cfg/TIMIT_baselines/TIMIT_liGRU_fmllr.cfg
, but confused by the three types of features.Shall I run the
run.sh
with three types of features three times and change paths likefea_lst
and`--utt2spk=? Should the
$KALDI-ROOT/egs/timit/s5/data/train/feats.scpdifferent across different feature types such as
$KALDI-ROOT/egs/timit/s5/data/train/feats.scpand
$KALDI-ROOT/egs/timit/s5/data-fmllr-tri3/train/feats.scp` ?