I had another question related to the WSJ S5 recipe. I understand that training of the tri2b (LDA+MLLT) system involves in training LDA transform which manifests itself as a file full.mat in the tri2b folder (generated by train_lda_mllt.sh).
The tri3b recipe then picks up from where the tri2b left off, with a call to train_sat.sh which aims to give an LDA+MLLT+SAT system. The problem I encounter next is with the the call to align_fmllr.sh in run.sh:
align_fmllr.sh complains that full.mat does not exist in tri3b. It will not exist since there was no direct call to train_lda_mllt.sh while training the models in the tri3b folder, is that right? Is there a bug somewhere, or were there some mis-steps in running the s5 recipe? Otherwise should I just be linking to the full.mat from the tri2b folder?
Last edit: aanchan 2014-07-02
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
You may be using an older version of the scripts. In the up-to-date trunk,
full.mat is copied by train_sat.sh into its own directory (e.g. tri3b). In
some older versions it may not have. Regardless of this, the warnings are
harmless.
Dan
I had another question related to the WSJ S5 recipe. I understand that
training of the tri2b (LDA+MLLT) system involves in training LDA transform
which manifests itself as a file full.mat in the tri2b folder (generated by
train_lda_mllt.sh).
The tri3b recipe then picks up from where the tri2b left off, with a call
to train_sat.sh which aims to give an LDA+MLLT+SAT system. The problem I
encounter next is with the the call to align_fmllr.sh in run.sh:
From 3b system, align all si284 data.
align_fmllr.sh complains that full.mat does not exist in tri3b. It will
not exist since there was no direct call to train_lda_mllt.sh while
training the models in the tri3b folder, is that right? Is there a bug
somewhere, or were there some mis-steps in running the s5 recipe? Otherwise
should I just be linking to the full.mat from the tri2b folder?
Hello,
I had another question related to the WSJ S5 recipe. I understand that training of the tri2b (LDA+MLLT) system involves in training LDA transform which manifests itself as a file full.mat in the tri2b folder (generated by train_lda_mllt.sh).
The tri3b recipe then picks up from where the tri2b left off, with a call to train_sat.sh which aims to give an LDA+MLLT+SAT system. The problem I encounter next is with the the call to align_fmllr.sh in run.sh:
--Comment-- From 3b system, align all si284 data.
steps/align_fmllr.sh --nj 20 --cmd "$train_cmd" data/train_si284 data/lang exp/tri3b exp/tri3b_ali_si284 || exit 1;
align_fmllr.sh complains that full.mat does not exist in tri3b. It will not exist since there was no direct call to train_lda_mllt.sh while training the models in the tri3b folder, is that right? Is there a bug somewhere, or were there some mis-steps in running the s5 recipe? Otherwise should I just be linking to the full.mat from the tri2b folder?
Last edit: aanchan 2014-07-02
You may be using an older version of the scripts. In the up-to-date trunk,
full.mat is copied by train_sat.sh into its own directory (e.g. tri3b). In
some older versions it may not have. Regardless of this, the warnings are
harmless.
Dan
On Wed, Jul 2, 2014 at 11:06 AM, aanchan aanchan-mohan@users.sf.net wrote: