Hi Nicolay,
The second error discussed in Assertion j==n_total_map failed topic, i.e
init_gau not producing gauden_counts for 1s_c feature sets still in place.
Running init_gau with parameters from log file just in terminal causes
"Segmentation fault" error message. Any ideas how to make it run?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Sure, which files do you need? .lda and .fileids used in parameters for sure
will be required and there will be no problem, but feat directory has about
800MB do yo need these files too?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The reason of the crash is that your lda matrix is very small (13x13) while
your ldadim is set to 29. I've just committed a fix where
it prints more meaningful message in this situation instead of crash
I suggest you to rebuild LDA matrix. Probably you somehow screwed the
configuration before it was trained. If you changed feat to 1s_c, you probably
wanted to train with cepwin. Then matrix size should be something like 39x29.
You can dump LDA matrix with printp command.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
you can take a look on my config in the message 14 of thread https://sourcefo
rge.net/projects/cmusphinx/forums/forum/382337/topic/3744788. this
configuration works fine for mllt and 1s_c for previous versions of ST. I
understand that some config options may have changed due to MMIE training
therefore I would be grateful if you could take a look at my config and
suggest what I have to set in order to get proper LDA matrix
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Your changes are ok, but probably you made a typo in CFG_FEAT_WINDOW.
Doublecheck in the logs that lda training binaries were invoked with -cepwin 3
option.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi Nicolay,
The second error discussed in Assertion j==n_total_map failed topic, i.e
init_gau not producing gauden_counts for 1s_c feature sets still in place.
Running init_gau with parameters from log file just in terminal causes
"Segmentation fault" error message. Any ideas how to make it run?
I think it's easy to fix that if you'll send me the files.
Sure, which files do you need? .lda and .fileids used in parameters for sure
will be required and there will be no problem, but feat directory has about
800MB do yo need these files too?
Try to run it yourself without feat files. If it still crashes, its easier
then.
It crashes on first mfc file so I'm sending you files on your email in few
minutes
The reason of the crash is that your lda matrix is very small (13x13) while
your ldadim is set to 29. I've just committed a fix where
it prints more meaningful message in this situation instead of crash
I suggest you to rebuild LDA matrix. Probably you somehow screwed the
configuration before it was trained. If you changed feat to 1s_c, you probably
wanted to train with cepwin. Then matrix size should be something like 39x29.
You can dump LDA matrix with printp command.
you can take a look on my config in the message 14 of thread https://sourcefo
rge.net/projects/cmusphinx/forums/forum/382337/topic/3744788. this
configuration works fine for mllt and 1s_c for previous versions of ST. I
understand that some config options may have changed due to MMIE training
therefore I would be grateful if you could take a look at my config and
suggest what I have to set in order to get proper LDA matrix
Your changes are ok, but probably you made a typo in CFG_FEAT_WINDOW.
Doublecheck in the logs that lda training binaries were invoked with -cepwin 3
option.
You're right, an extra character in CFG_FEAT_WINDOW name caused -cepwin 0.
Thanks a lot