Help on Sphinx3 Training and Decoding

[Senjam Shantirani]

Query 1: PLEASE COMMENT IF MY TRAING IS FINE BASE ON THE POINTS a) & b) MENTIONED BELOW.

a) Training ended like shown below,

---

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 2 ERROR messages and 0 WARNING messages. Please check the log file for details.
Normalization for iteration: 7
Current Overall Likelihood Per Frame = 1.92801139707547
Split Gaussians, increase by 0
Training for 8 Gaussian(s) completed after 7 iterations
MODULE: 90 deleted interpolation
Skipped for continuous models
MODULE: 99 Convert to Sphinx2 format models
Can not create models used by Sphinx-II.
If you intend to create models to use with Sphinx-II models, please rerun with:
$ST::CFG_HMM_TYPE = '.semi.' or
$ST::CFG_HMM_TYPE = '.cont' and $ST::CFG_FEATURE = '1s_12c_12d_3p_12dd' and $ST::CFG_STATESPERHMM = '5'
root@shanti-Satellite-C650:/home/PhoneModel/workspace#

---

b) I got 20 folders in my model_parameters, each having the files mean, variance, transition_matrices and mixture weight.

timit.cd_cont_1000
timit.cd_cont_1000_1
timit.cd_cont_1000_2
timit.cd_cont_1000_4
timit.cd_cont_1000_8
timit.cd_cont_initial
timit.cd_cont_untied
timit.ci_cont
timit.ci_cont_1
timit.ci_cont_2
timit.ci_cont_4
timit.ci_cont_8
timit.ci_cont_flatinitial
timit.ci_cont_initial
timit.falign_ci_cont
timit.falign_ci_cont_1
timit.falign_ci_cont_2
timit.falign_ci_cont_4
timit.falign_ci_cont_8
timit.falign_ci_cont_initial

Query 2: WITH THE ABOVE TRAINED DATA, I TRIED TO DO DECODING, BUT NOT SURE WHAT TO MODIFY IN THE

This variables, used by the decoder, have to be user defined, and

may affect the decoder output

$DEC_CFG_LANGUAGEMODEL_DIR = "$DEC_CFG_BASE_DIR/etc";
$DEC_CFG_LANGUAGEMODEL = "$DEC_CFG_LANGUAGEMODEL_DIR/timit.lm.DMP"; (WHAT TO WRITE HERE)
$DEC_CFG_LANGUAGEWEIGHT = "10";
$DEC_CFG_BEAMWIDTH = "1e-80";
$DEC_CFG_WORDBEAM = "1e-40";

But when I ran the command "perl scripts_pl/decode/slave.pl ", the error came like shown below:

root@shanti-Satellite-C650:/home/PhoneModel/workspace# perl scripts_pl/decode/slave.pl
MODULE: DECODE Decoding using models previously trained
Decoding 4620 segments starting at 0 (part 1 of 1)
0%
This step had 2 ERROR messages and 7 WARNING messages. Please check the log file for details.
Aligning results to find error rate
Can't open /home/PhoneModel/workspace/result/timit-1-1.match
word_align.pl failed with error code 65280 at scripts_pl/decode/slave.pl line 172.

In the logdir/decode, the error reported is:

INFO: cont_mgau.c(510): Reading mixture weights file '/home/PhoneModel/workspace/model_parameters/timit.cd_cont_1000/mixture_weights'
ERROR: "cont_mgau.c", line 653: Weight normalization failed for 12 senones
INFO: cont_mgau.c(665): Read 1132 x 15 mixture weights
INFO: cont_mgau.c(693): Removing uninitialized Gaussian densities
9 10 11 48 49 50 79 90 91 92 263 264 265 510 511 512 790 844 845 846
WARNING: "cont_mgau.c", line 767: 464 densities removed (20 mixtures removed entirely)
INFO: cont_mgau.c(783): Applying variance floor
INFO: cont_mgau.c(801): 613 variance values floored
INFO: cont_mgau.c(849): Precomputing Mahalanobis distance invariants
INFO: tmat.c(169): Reading HMM transition probability matrices: /home/PhoneModel/workspace/model_parameters/timit.cd_cont_1000/transition_matrices
WARNING: "tmat.c", line 242: Normalization failed for tmat 16 from state 0
WARNING: "tmat.c", line 242: Normalization failed for tmat 16 from state 1
WARNING: "tmat.c", line 242: Normalization failed for tmat 16 from state 2
WARNING: "tmat.c", line 242: Normalization failed for tmat 30 from state 0
WARNING: "tmat.c", line 242: Normalization failed for tmat 30 from state 1
WARNING: "tmat.c", line 242: Normalization failed for tmat 30 from state 2
INFO: Initialization of tmat_t, report:
INFO: Read 44 transition matrices of size 3x4
INFO:
INFO: dict.c(475): Reading main dictionary: /home/PhoneModel/workspace/etc/timit.dic
INFO: dict.c(478): 133011 words read
INFO: dict.c(483): Reading filler dictionary: /home/PhoneModel/workspace/etc/timit.filler
INFO: dict.c(486): 3 words read
INFO: Initialization of dict_t, report:
INFO: No of CI phone: 0
INFO: Max word: 137110
INFO: No of word: 133014
INFO:
INFO: lm.c(606): LM read('/home/PhoneModel/workspace/etc/timit.lm.DMP', lw= 10.00, wip= 0.20, uw= 0.70)
INFO: lm.c(608): Reading LM file /home/PhoneModel/workspace/etc/timit.lm.DMP (LM name "default")
SYSTEM_ERROR: "lm_3g_dmp.c", line 1270: fopen(/home/PhoneModel/workspace/etc/timit.lm.DMP,rb) failed
; No such file or directory
Sun Mar 6 21:00:32 2016

I cant find the file /result/timit-1-1.match in the result folder.

Thanking you,
Shanti

[Senjam Shantirani]

After digging further, I learn the issue is absence if Language Model, I created one uisng the LMTOOL

http://www.speech.cs.cmu.edu/tools/lmtool-new.html

and modified the sphinx_decode.cfg with the new LM:
$DEC_CFG_LANGUAGEMODEL = "$DEC_CFG_LANGUAGEMODEL_DIR/timit.lm";

Then, execution gave rise to the generation of timit-1-1.match in the result folder, in the main working folder and finally got the WER and SER.

But I have a query if the LM I created is using Online Link mentioned above is okay to use for all cases.

Regards,
Shanti

[Nickolay V. Shmyrev]

But I have a query if the LM I created is using Online Link mentioned above is okay to use for all cases.

It is fine.

[Senjam Shantirani]

I guess my training is a sucessful one as per the terminal output and the files obtained in my model_parameter folder. Please Comment.

And thank you for your assistance.