Train errors

[wilson.tang]

hello，
Hello
When i'm using the parameters
An error occurred during training：
If I open the LDA parameters,
$CFG_LDA_MLLT = 'yes';
$CFG_LDA_DIMENSION = 32;

And open:
$CFG_FULLVAR = 'yes';
$CFG_DIAGFULL = 'yes';

The following error occurred:

INFO: feat.c(715): Initializing feature stream to type: '1s_c_d_dd', ceplen=13, CMN='batch', VARNORM='yes', AGC='none'
INFO: main.c(318): Reading E:/pocketsphinx/train-build/thchs20-phone-16K/model_architecture/thchs20-phone-16K.ci.mdef
INFO: model_def_io.c(573): Model definition info:
INFO: model_def_io.c(574): 67 total models defined (67 base, 0 tri)
INFO: model_def_io.c(575): 268 total states
INFO: model_def_io.c(576): 201 total tied states
INFO: model_def_io.c(577): 201 total tied CI states
INFO: model_def_io.c(578): 67 total tied transition matrices
INFO: model_def_io.c(579): 4 max state/model
INFO: model_def_io.c(580): 4 min state/model
INFO: s3mixw_io.c(117): Read E:/pocketsphinx/train-build/thchs20-phone-16K/model_parameters/thchs20-phone-16K.ci_lda_flatinitial/mixture_weights [201x1x1 array]
INFO: s3tmat_io.c(118): Read E:/pocketsphinx/train-build/thchs20-phone-16K/model_parameters/thchs20-phone-16K.ci_lda_flatinitial/transition_matrices [67x3x4 array]
INFO: mod_inv.c(301): inserting tprob floor 1.000000e-05 and renormalizing
INFO: s3gau_io.c(169): Read E:/pocketsphinx/train-build/thchs20-phone-16K/model_parameters/thchs20-phone-16K.ci_lda_flatinitial/means [201x1x1 array]
ERROR: "s3gau_io.c", line 129: Failed to read parameter file E:/pocketsphinx/train-build/thchs20-phone-16K/model_parameters/thchs20-phone-16K.ci_lda_flatinitial/variances (expected 7839 values, got 305721)
FATAL: "main.c", line 1839: initialization failed
Sat Feb 18 11:38:27 2017

[wilson.tang]

Only when I open the LDA option will appear
$CFG_LDA_MLLT = 'yes';
$CFG_LDA_DIMENSION = 32;
And
$CFG_FULLVAR = 'yes';
$CFG_DIAGFULL = 'yes';

The two are not open at the same time, not mistakes, their training are normal

[Nickolay V. Shmyrev]

Fullvar and MLLT are not compatible. MLLT is an approach to compensate DIAG covariance matrices. So FULLVAR does not make much sense together with MLLT.

[wilson.tang]

Okay thank you.
I want to find Fullvar and MLLT performance differences,
But it seems that the performance difference between the two is not very obvious
For 67 phonemes, about 30 hours of training voice

[Nickolay V. Shmyrev]

MLLT should approach fullvar in terms of accuracy, but should be much faster and efficient.

[wilson.tang]

thank you very much for your reply