Help to improve accuracy

[Shammur]

Hello,

I am training a continuous model with around 20K sentences, which has around
98785 n_tri using cmu dic

my LM has perplexity on test text around 14.3

I have trained my last model with 3 states/hmm , 32 GMixture, and 3000 tied
states
I used forced Alignment .
my wav files are in sphere nist format

please do let me know what can I do to improve it, at the present my wer =
91.7% which is really not expected.
I have tried with diff tied states but dont know which one I used use or is
there any other calculation by which I can compute it.

my model , etc , logdir a sample wav and align file is in the link below
http://www.mediafire.com/file/5qy1x01186tqzdr/help.tar.gz

Please any suggestion or any sort of comment will be really helpful.

At this moment I am re running the training with 3s/hmm 16 gm and 8000 tied
states. which takes around 15hrs to train

Thanks in advance

[Shammur]

And here is other details:

Mon Apr 23 09:10:55 CEST 2012
MODULE: 00 verify training files
O.S. is case sensitive ("A" != "a").
Phones will be treated as case sensitive.
Phase 1: DICT - Checking to see if the dict and filler dict agrees with the
phonelist file.
Found 133120 words using 40 phones
Phase 2: DICT - Checking to make sure there are not duplicate entries in the
dictionary
Phase 3: CTL - Check general format; utterance length (must be positive);
files exist
Phase 4: CTL - Checking number of lines in the transcript should match lines
in control file
Phase 5: CTL - Determine amount of training data, see if n_tied_states seems
reasonable.
Estimated Total Hours Training: 15.6430972222222
Rule of thumb suggests 3000, however there is no correct answer
Phase 6: TRANSCRIPT - Checking that all the words in the transcript are in the
dictionary
Words in dictionary: 133117
Words in filler dictionary: 3
Phase 7: TRANSCRIPT - Checking that all the phones in the transcript are in
the phonelist, and all phones in the phonelist appear at least once
MODULE: 01 Train LDA transformation
Skipped (set $CFG_LDA_MLLT = 'yes' to enable)
MODULE: 02 Train MLLT transformation
Skipped (set $CFG_LDA_MLLT = 'yes' to enable)
MODULE: 05 Vector Quantization
Skipped for continuous models
MODULE: 10 Training Context Independent models for forced alignment and VTLN
Phase 1: Cleaning up directories:
accumulator...logs...qmanager...models...
Phase 2: Flat initialize
Phase 3: Forward-Backward
Baum welch starting for 1 Gaussian(s), iteration: 1 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 30 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 1
Current Overall Likelihood Per Frame = 26.1319655403433
Baum welch starting for 1 Gaussian(s), iteration: 2 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 12 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 2
Current Overall Likelihood Per Frame = 26.2070756188023
Training completed after 2 iterations
MODULE: 11 Force-aligning transcripts
Phase 1: Cleaning up directories:
logs...output...qmanager...
Phase 3: Creating dictionary for alignment...
Phase 4: Creating transcript for alignment...
Phase 5: Running force alignment in 1 parts
Force alignment starting: (1 of 1)
0%
MODULE: 12 Force-aligning data for VTLN
Skipped: $ST::CFG_VTLN set to 'no' in sphinx_train.cfg
MODULE: 20 Training Context Independent models
Phase 1: Cleaning up directories:
accumulator...logs...qmanager...models...
Phase 2: Copy initialize from falign model
Phase 3: Forward-Backward
Baum welch starting for 1 Gaussian(s), iteration: 1 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Normalization for iteration: 1
Current Overall Likelihood Per Frame = 26.2272230474393
Baum welch starting for 1 Gaussian(s), iteration: 2 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Normalization for iteration: 2
Current Overall Likelihood Per Frame = 26.3358971786455
Convergence Ratio = 0.108674131206236
Baum welch starting for 1 Gaussian(s), iteration: 3 (1 of 1)
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: 3
Current Overall Likelihood Per Frame = 26.5023794074265
Convergence Ratio = 0.166482228781007
Baum welch starting for 1 Gaussian(s), iteration: 4 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 18 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 4
Current Overall Likelihood Per Frame = 26.7898282866141
Convergence Ratio = 0.2874488791876
Baum welch starting for 1 Gaussian(s), iteration: 5 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 96 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 5
Current Overall Likelihood Per Frame = 26.9731678517901
Convergence Ratio = 0.183339565176016
Baum welch starting for 1 Gaussian(s), iteration: 6 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 168 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 6
Current Overall Likelihood Per Frame = 27.0386115539069
Training completed after 6 iterations
MODULE: 30 Training Context Dependent models
Phase 1: Cleaning up directories:
accumulator...logs...qmanager...
Phase 2: Initialization
Phase 3: Forward-Backward
Baum welch starting for iteration: 1 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 204 ERROR messages and 1 WARNING messages. Please check the log
file for details.
Normalization for iteration: 1
WARNING: This step had 0 ERROR messages and 96 WARNING messages. Please check
the log file for details.
Current Overall Likelihood Per Frame = 27.0620323767413
Baum welch starting for iteration: 2 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 200 ERROR messages and 1 WARNING messages. Please check the log
file for details.
Normalization for iteration: 2
WARNING: This step had 0 ERROR messages and 96 WARNING messages. Please check
the log file for details.
Current Overall Likelihood Per Frame = 27.3165607504742
Training completed after 2 iterations
MODULE: 40 Build Trees
Phase 1: Cleaning up old log files...
Phase 2: Make Questions
Phase 3: Tree building
Processing each phone with each state
AA 0
AA 1
AA 2
AE 0
AE 1
AE 2
AH 0
AH 1
AH 2
AO 0
AO 1
AO 2
AW 0
AW 1
AW 2
AY 0
AY 1
AY 2
B 0
B 1
B 2
CH 0
CH 1
CH 2
D 0
D 1
D 2
DH 0
DH 1
DH 2
EH 0
EH 1
EH 2
ER 0
ER 1
ER 2
EY 0
EY 1
EY 2
F 0
F 1
F 2
G 0
G 1
G 2
HH 0
HH 1
HH 2
IH 0
IH 1
IH 2
IY 0
IY 1
IY 2
JH 0
JH 1
JH 2
K 0
K 1
K 2
L 0
L 1
L 2
M 0
M 1
M 2
N 0
N 1
N 2
NG 0
NG 1
NG 2
OW 0
OW 1
OW 2
OY 0
OY 1
OY 2
P 0
P 1
P 2
R 0
R 1
R 2
S 0
S 1
S 2
SH 0
SH 1
SH 2
T 0
T 1
T 2
TH 0
TH 1
TH 2
UH 0
UH 1
UH 2
UW 0
UW 1
UW 2
V 0
V 1
V 2
W 0
W 1
W 2
Y 0
Y 1
Y 2
Z 0
Z 1
Z 2
ZH 0
ZH 1
ZH 2
Skipping SIL
MODULE: 45 Prune Trees
Phase 1: Tree Pruning
Phase 2: State Tying
MODULE: 50 Training Context dependent models
Phase 1: Cleaning up directories:
accumulator...logs...qmanager...
Phase 2: Copy CI to CD initialize
Phase 3: Forward-Backward
Baum welch starting for 1 Gaussian(s), iteration: 1 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 204 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 1
This step had 3 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Current Overall Likelihood Per Frame = 27.0620323767413
Baum welch starting for 1 Gaussian(s), iteration: 2 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 200 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 2
Current Overall Likelihood Per Frame = 27.1756667388801
Convergence Ratio = 0.113634362138786
Baum welch starting for 1 Gaussian(s), iteration: 3 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 210 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 3
Current Overall Likelihood Per Frame = 27.2068794897292
Split Gaussians, increase by 1
Current Overall Likelihood Per Frame = 27.2068794897292
Convergence Ratio = 0.0312127508491464
Baum welch starting for 2 Gaussian(s), iteration: 1 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 208 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 1
This step had 147 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Current Overall Likelihood Per Frame = 27.0087931553322
Baum welch starting for 2 Gaussian(s), iteration: 2 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 210 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 2
Current Overall Likelihood Per Frame = 27.2304503029369
Convergence Ratio = 0.221657147604699
Baum welch starting for 2 Gaussian(s), iteration: 3 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 208 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 3
Current Overall Likelihood Per Frame = 27.2678195942578
Split Gaussians, increase by 2
Current Overall Likelihood Per Frame = 27.2678195942578
Convergence Ratio = 0.0373692913208608
Baum welch starting for 4 Gaussian(s), iteration: 1 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 206 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 1
This step had 316 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Current Overall Likelihood Per Frame = 27.0696735549779
Baum welch starting for 4 Gaussian(s), iteration: 2 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 208 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 2
Current Overall Likelihood Per Frame = 27.3343836749616
Convergence Ratio = 0.264710119983711
Baum welch starting for 4 Gaussian(s), iteration: 3 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 206 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 3
Current Overall Likelihood Per Frame = 27.4163755213446
Split Gaussians, increase by 4
Current Overall Likelihood Per Frame = 27.4163755213446
Convergence Ratio = 0.0819918463830192
Baum welch starting for 8 Gaussian(s), iteration: 1 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 210 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 1
This step had 516 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Current Overall Likelihood Per Frame = 27.2537423845428
Baum welch starting for 8 Gaussian(s), iteration: 2 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 208 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 2
Current Overall Likelihood Per Frame = 27.532755561439
Convergence Ratio = 0.279013176896225
Baum welch starting for 8 Gaussian(s), iteration: 3 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 208 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 3
Current Overall Likelihood Per Frame = 27.6295347126804
Split Gaussians, increase by 8
Current Overall Likelihood Per Frame = 27.6295347126804
Convergence Ratio = 0.0967791512413747
Baum welch starting for 16 Gaussian(s), iteration: 1 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 206 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 1
This step had 1152 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Current Overall Likelihood Per Frame = 27.4939123333241
Baum welch starting for 16 Gaussian(s), iteration: 2 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 204 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 2
Current Overall Likelihood Per Frame = 27.8240073740902
Convergence Ratio = 0.330095040766142
Baum welch starting for 16 Gaussian(s), iteration: 3 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 200 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 3
Current Overall Likelihood Per Frame = 27.9680046230904
Convergence Ratio = 0.143997249000204
Baum welch starting for 16 Gaussian(s), iteration: 4 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 200 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 4
Current Overall Likelihood Per Frame = 28.0956212601184
Convergence Ratio = 0.127616637028019
Baum welch starting for 16 Gaussian(s), iteration: 5 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 200 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 5
Current Overall Likelihood Per Frame = 28.2003838146574
Convergence Ratio = 0.10476255453899
Baum welch starting for 16 Gaussian(s), iteration: 6 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 196 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 6
Current Overall Likelihood Per Frame = 28.2845113158031
Split Gaussians, increase by 16
Current Overall Likelihood Per Frame = 28.2845113158031
Convergence Ratio = 0.0841275011457157
Baum welch starting for 32 Gaussian(s), iteration: 1 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 196 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 1
This step had 1838 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Current Overall Likelihood Per Frame = 28.1444890706236
Baum welch starting for 32 Gaussian(s), iteration: 2 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 194 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 2
Current Overall Likelihood Per Frame = 28.4805715265889
Convergence Ratio = 0.336082455965343
Baum welch starting for 32 Gaussian(s), iteration: 3 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 196 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 3
Current Overall Likelihood Per Frame = 28.623325488486
Convergence Ratio = 0.142753961897132
Baum welch starting for 32 Gaussian(s), iteration: 4 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 196 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 4
Current Overall Likelihood Per Frame = 28.7337483050926
Convergence Ratio = 0.110422816606597
Baum welch starting for 32 Gaussian(s), iteration: 5 (1 of 1)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
This step had 196 ERROR messages and 0 WARNING messages. Please check the log
file for details.
Normalization for iteration: 5
Current Overall Likelihood Per Frame = 28.8125174122122
Split Gaussians, increase by 0
Training for 32 Gaussian(s) completed after 5 iterations
MODULE: 60 Lattice Generation
Skipped: $ST::CFG_MMIE set to 'no' in sphinx_train.cfg
MODULE: 61 Lattice Pruning
Skipped: $ST::CFG_MMIE set to 'no' in sphinx_train.cfg
MODULE: 62 Lattice Format Conversion
Skipped: $ST::CFG_MMIE set to 'no' in sphinx_train.cfg
MODULE: 65 MMIE Training
Skipped: $ST::CFG_MMIE set to 'no' in sphinx_train.cfg
MODULE: 90 deleted interpolation
Skipped for continuous models
Tue Apr 24 03:55:36 CEST 2012

[Nickolay V. Shmyrev]

Your feature extraction is not correct

You wav files are encoded with shorten.

You need to decode them with sph2pipe before training or you can use
-sph2pipe yes option of sphinx_fe. You need to have sph2pipe anyway. To
modify the option you need to edit the feature extraction script, there is no
way to do it from a configuration file.

[Shammur]

Thank You
I am trying now and let you know the result

[Shammur]

Hello,

I have tried sph2pipe option.
my new model: http://www.mediafire.com/file/lrdepc4aqk2ndey/help_2.tar.gz
Still I am getting a wer of 92.5%

Is my feature extraction ok ?
or this error is due to choice of wrong tied state

Any way to decide what tied state should I use?

His training has parameters: 3 s/hmm, 16 gmix, tied state = 8000

Please do suggest me

Thanks in advance

[Nickolay V. Shmyrev]

It doesn't seem you used sph2pipe right. You need to provide a feature file
example in order to let me say more. You can control CMN value in the logs, it
must be around 10-12.

[Shammur]

Here is one example:
http://www.mediafire.com/file/vzuchao2173fcbe/0k0011ss.wav
http://www.mediafire.com/file/237y2ut5gj21uug/0k0011ss.mfc

Please do let me know if any files will help you to suggest me

[Nickolay V. Shmyrev]

mfc file is not correct. It doesn't sound like you installed sph2pipe
properly. You can dump it with

 sphinx_cepview -f file.mfc

The proper values should look like:

  6.951  -0.736  -0.112  -0.182  -0.146  -0.017  -0.261  -0.116  -0.202  -0.078 
  6.483  -0.736  -0.040  -0.243  -0.100  -0.072  -0.281  -0.076  -0.079  -0.008 
  6.676  -0.795  -0.136  -0.213  -0.112  -0.061  -0.325   0.039   0.006  -0.065 
  6.898  -0.614  -0.293  -0.249  -0.145  -0.193  -0.139  -0.042  -0.116  -0.078 
  7.093  -0.675  -0.152  -0.284  -0.194  -0.129  -0.317  -0.076  -0.081  -0.071 
  6.938  -0.567  -0.056  -0.235  -0.159  -0.054  -0.358  -0.211  -0.063  -0.048

[Shammur]

Thank You,

It works now, WER =8.3%

Now will be checking how the results differ with different parameters

Thanks for all you help

[Shammur]

Can you please give me any information about how can I control CMN value .

Thanks

[Nickolay V. Shmyrev]

Can you please give me any information about how can I control CMN value .

When I wrote you this:

You can control CMN value in the logs, it must be around 10-12.

I meant that you can open the log and read it and see the first value in the
CMN string.
It looks like this

INFO: cmn.c(175): CMN:  7.30 -0.08 -0.06  0.19 -0.33 -0.00 -0.12 -0.01  0.07  0.01 -0.21 -0.18 -0.

You need to make sure that the first value is reasonable.