Please help me if you can: New feature with 6 dimensions in Kaldi

[hang phuong Nguyen]

Hi everyone.
I am a new kaldi. Now I am doing a new problem in Kaldi. I created new features which had 6 dimensions, then replaced MFCC features by my features and built a training with monophone. But after training process, I got WER = 99.98% and I havenot found the solution.
I did 40 interactions in training process and this are contents in acc.39.1.log, align.38.1.log, update.39.log.
Please give me some comments about what is my error and help me if you can. Thank you so much.

acc.39.1.log:
# gmm-acc-stats-ali exp/mono6/39.mdl "ark,s,cs:apply-cmvn  --utt2spk=ark:data/train/split2/1/utt2spk scp:data/train/split2/1/cmvn.scp scp:data/train/split2/1/feats.scp ark:- | add-deltas --delta-order=0 ark:- ark:- |" "ark:gunzip -c exp/mono6/ali.1.gz|" exp/mono6/39.1.acc
# Started at Tue Dec 20 11:57:13 ICT 2016
#
gmm-acc-stats-ali exp/mono6/39.mdl 'ark,s,cs:apply-cmvn  --utt2spk=ark:data/train/split2/1/utt2spk scp:data/train/split2/1/cmvn.scp scp:data/train/split2/1/feats.scp ark:- | add-deltas --delta-order=0 ark:- ark:- |' 'ark:gunzip -c exp/mono6/ali.1.gz|' exp/mono6/39.1.acc
apply-cmvn --utt2spk=ark:data/train/split2/1/utt2spk scp:data/train/split2/1/cmvn.scp scp:data/train/split2/1/feats.scp ark:-
add-deltas --delta-order=0 ark:- ark:-
LOG (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:105) Processed 50 utterances; for utterance F00-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F00_Dialog170_MIC01 avg. like is -31.6046 over 267 frames.
LOG (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:105) Processed 100 utterances; for utterance F00-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F00_News110_MIC01 avg. like is -31.3081 over 769 frames.
LOG (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:105) Processed 150 utterances; for utterance F01-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F01_Dialog250_MIC01 avg. like is -31.5624 over 242 frames
....
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP850067F30VPH2
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP860068F30VPH2
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP880070F30VPH1
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP890071F30VPH1
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP900072F30VPH1
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP900072F30VPH2
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP910073F30VPH1
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP910073F30VPH2
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP920074F30VPH2
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP930075F30VPH2
WARNING (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:79) No alignment for utterance FPH-PHP940076F30VPH1
LOG (apply-cmvn:main():apply-cmvn.cc:146) Applied cepstral mean normalization to 1454 utterances, errors on 0
LOG (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:112) Done 1266 files, 188 with errors.
LOG (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:115) Overall avg like per frame (Gaussian only) = -31.3677 over 1395804 frames.
LOG (gmm-acc-stats-ali:main():gmm-acc-stats-ali.cc:123) Written accs.
# Accounting: time=4 threads=1
# Ended (code 0) at Tue Dec 20 11:57:17 ICT 2016, elapsed time 4 seconds


align.38.1.log:
# gmm-align-compiled --transition-scale=1.0 --acoustic-scale=0.1 --self-loop-scale=0.1 --beam=15 --retry-beam=60 --careful=false "gmm-boost-silence --boost=1.0 1 exp/mono6/38.mdl - |" "ark:gunzip -c exp/mono6/fsts.1.gz|" "ark,s,cs:apply-cmvn  --utt2spk=ark:data/train/split2/1/utt2spk scp:data/train/split2/1/cmvn.scp scp:data/train/split2/1/feats.scp ark:- | add-deltas --delta-order=0 ark:- ark:- |" "ark,t:|gzip -c >exp/mono6/ali.1.gz"
# Started at Tue Dec 20 10:58:25 ICT 2016
#
gmm-align-compiled --transition-scale=1.0 --acoustic-scale=0.1 --self-loop-scale=0.1 --beam=15 --retry-beam=60 --careful=false 'gmm-boost-silence --boost=1.0 1 exp/mono6/38.mdl - |' 'ark:gunzip -c exp/mono6/fsts.1.gz|' 'ark,s,cs:apply-cmvn  --utt2spk=ark:data/train/split2/1/utt2spk scp:data/train/split2/1/cmvn.scp scp:data/train/split2/1/feats.scp ark:- | add-deltas --delta-order=0 ark:- ark:- |' 'ark,t:|gzip -c >exp/mono6/ali.1.gz'
gmm-boost-silence --boost=1.0 1 exp/mono6/38.mdl -
WARNING (gmm-boost-silence:main():gmm-boost-silence.cc:82) The pdfs for the silence phones may be shared by other phones (note: this probably does not matter.)
LOG (gmm-boost-silence:main():gmm-boost-silence.cc:93) Boosted weights for 5 pdfs, by factor of 1
LOG (gmm-boost-silence:main():gmm-boost-silence.cc:103) Wrote model to -
add-deltas --delta-order=0 ark:- ark:-
apply-cmvn --utt2spk=ark:data/train/split2/1/utt2spk scp:data/train/split2/1/cmvn.scp scp:data/train/split2/1/feats.scp ark:-
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance F00-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F00_General65_MIC01 with beam 60
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance F00-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F00_General67_MIC01 with beam 60
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance F00-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F00_General68_MIC01 with beam 60
....
LOG (apply-cmvn:main():apply-cmvn.cc:146) Applied cepstral mean normalization to 1454 utterances, errors on 0
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance FPH-PHP940076F30VPH1 with beam 60
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:475) Did not successfully decode file FPH-PHP940076F30VPH1, len = 1809
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance FPH-PHP940076F30VPH2 with beam 60
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance FPH-PHP950077F30VPH1 with beam 60
WARNING (gmm-align-compiled:AlignUtteranceWrapper():decoder-wrappers.cc:466) Retrying utterance FPH-PHP950077F30VPH2 with beam 60
LOG (gmm-align-compiled:main():gmm-align-compiled.cc:129) Overall log-likelihood per frame is -31.8806 over 1395804 frames.
LOG (gmm-align-compiled:main():gmm-align-compiled.cc:131) Retried 919 out of 1454 utterances.
LOG (gmm-align-compiled:main():gmm-align-compiled.cc:133) Done 1266, errors on 188
# Accounting: time=3108 threads=1
# Ended (code 0) at Tue Dec 20 11:50:13 ICT 2016, elapsed time 3108 seconds


update.39.log:
# gmm-est --write-occs=exp/mono6/40.occs --mix-up=983 --power=0.25 exp/mono6/39.mdl "gmm-sum-accs - exp/mono6/39.*.acc|" exp/mono6/40.mdl
# Started at Tue Dec 20 11:57:17 ICT 2016
#
gmm-est --write-occs=exp/mono6/40.occs --mix-up=983 --power=0.25 exp/mono6/39.mdl 'gmm-sum-accs - exp/mono6/39.*.acc|' exp/mono6/40.mdl
gmm-sum-accs - exp/mono6/39.1.acc exp/mono6/39.2.acc
LOG (gmm-sum-accs:main():gmm-sum-accs.cc:63) Summed 2 stats, total count 2.96446e+06, avg like/frame -31.3633
LOG (gmm-sum-accs:main():gmm-sum-accs.cc:66) Total count of stats is 2.96446e+06
LOG (gmm-sum-accs:main():gmm-sum-accs.cc:67) Written stats to -
LOG (gmm-est:MleUpdate():transition-model.cc:393) TransitionModel::Update, objf change is 0 per frame over 2.96446e+06 frames.
LOG (gmm-est:MleUpdate():transition-model.cc:396) 104 probabilities floored, 364 out of 577 transition-states skipped due to insuffient data (it is normal to have some skipped.)
LOG (gmm-est:main():gmm-est.cc:102) Transition model update: Overall 0 log-like improvement per frame over 2.96446e+06 frames.
WARNING (gmm-est:MleDiagGmmUpdate():mle-diag-gmm.cc:365) Gaussian has too little data but not removing it because it is the last Gaussian: i = 0, occ = 0, weight = 1
WARNING (gmm-est:MleDiagGmmUpdate():mle-diag-gmm.cc:365) Gaussian has too little data but not removing it because it is the last Gaussian: i = 0, occ = 0, weight = 1
LOG (gmm-est:MleAmDiagGmmUpdate():mle-am-diag-gmm.cc:225) 0 variance elements floored in 0 Gaussians, out of 1027
LOG (gmm-est:MleAmDiagGmmUpdate():mle-am-diag-gmm.cc:229) Removed 0 Gaussians due to counts < --min-gaussian-occupancy=10 and --remove-low-count-gaussians=true
LOG (gmm-est:main():gmm-est.cc:113) GMM update: Overall 0.00956833 objective function improvement per frame over 2.96446e+06 frames
LOG (gmm-est:main():gmm-est.cc:116) GMM update: Overall avg like per frame = -31.3633 over 2.96446e+06 frames.
LOG (gmm-est:SplitByCount():am-diag-gmm.cc:116) Split 143 states with target = 983, power = 0.25, perturb_factor = 0.01 and min_count = 20, split #Gauss from 1027 to 1027
LOG (gmm-est:main():gmm-est.cc:146) Written model to exp/mono6/40.mdl
# Accounting: time=0 threads=1
# Ended (code 0) at Tue Dec 20 11:57:17 ICT 2016, elapsed time 0 seconds

[Daniel Povey]

The features aren't very good and the alignments are failing. It's
not that surprising. Maybe try adding your features to normal
features.

Dan


On Thu, Dec 29, 2016 at 12:37 AM, hang phuong Nguyen
<hangph...@gmail.com> wrote:
>
>
> Vào 15:23:21 UTC+7 Thứ Năm, ngày 29 tháng 12 năm 2016, hang phuong Nguyen đã
> viết:

> --
> You received this message because you are subscribed to the Google Groups
> "kaldi-help" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to kaldi-help+...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[hang phuong Nguyen]

Dear Mr Povey
Thank you for your reply.
In my work, I have to use just new features which have 6 dimensions. So I cannot add my features to normal features. I havenot had any ideas to solve these errors, Could you give me some advices? 
I can describe my process and Could you give me some comments?:
I computed my features by python and then saved in text file with the following format:
F00-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F00_Dialog121_MIC01  [
 -11.48139  -17.36911  -84.69856   38.74913   59.22892  -47.63833
 -87.09810   -4.06405   85.81645   55.20061  -65.87661   -0.95802
  84.66265   49.85245  -56.40939   11.76371  -55.68209    9.85124
 -69.52729  -51.58050  -10.77374   72.43873   29.38069  -41.37265
  18.54256  -85.50710  -23.65084  -70.13355  -62.93958    5.72132
  73.06228   49.34236  -55.47288  -46.45543   50.47929    8.08242
 -56.54454  -31.75079  -48.03521  -66.56364   23.58067   43.79872
 -48.03805    7.21895   50.22250   63.91903   35.73417  -76.13044
....
]

 then I convert to ark file and scp file with these commands:

## train

copy-feats --compress=true ark,t:/home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/digits_audio/audio/sscf/sscf_train.txt ark,scp:/home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/digits_audio/audio/sscf/sscf_train.ark,/home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/digits_audio/audio/sscf/sscf_train.scp

## test

copy-feats --compress=true ark,t:/home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/digits_audio/audio/sscf/sscf_test.txt ark,scp:/home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/digits_audio/audio/sscf/sscf_test.ark,/home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/digits_audio/audio/sscf/sscf_test.scp

then I computed the cmvn:

## train

steps/compute_cmvn_stats.sh /home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/data/train /home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/exp/make_sscf/train /home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/digits_audio/audio/sscf || exit 1;

## test

steps/compute_cmvn_stats.sh /home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/data/test /home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/exp/make_sscf/test /home/mica/MyKaldi/kaldi-trunk/egs/demo_sscf/digits_audio/audio/sscf || exit 1;

Finally, I replace MFCC by my features and built monophone training.

Thank you so much and I look forward to hearing from you.

[Daniel Povey]

I meant append, not add as in "+", see steps/append_feats.sh

[hang phuong Nguyen]

Thank you, Mr Povey.
I will see carefully steps/append_feats.sh

[hang phuong Nguyen]

Dear Mr Dan
Following your advice as using steps/append_feats.sh. But I do not understand: in my work, I have to use 6-dimensions features but in append_feats.sh, it ask that have at least two data sources to combine. So I can not use append_feats.sh.

Please give me some advices.

Thank you

[Daniel Povey]

I was imagining that this is a speech recognition task and you would be appending your new 6-dimensional features with the regular MFCC features.

If that's really all the features you have, you could try increasing the beams in the alignment (e.g. change the beam=6 in train_mono.sh to beam=12).

[hang phuong Nguyen]

Thank you so much for your advice.
I would try doing it.

Hang Phuong

[hang phuong Nguyen]

And If you have free time, you could explain to me: What is the "beam" in Kaldi, its use and its influence in the alignment.I saw it and read all of code which related with "beam". But I have not been clearly.

Thank you so much.

Hang Phuong

 

[Daniel Povey]

it's the pruning beam of Viterbi decoding with beam-search- you can use those for search terms but I don't have time to explain more.

[hang phuong Nguyen]

Thank you so much, Mr Dan

[hang phuong Nguyen]

Dear Mr Dan,
I did train monophone with beam = 12 as your advice. And then i decoded with config:
first_beam=10.0
beam=15.0
lattice_beam=12.0

 But I saw that decode was failed, the program was not decode with all of wav file in test database and the content in decode.log is
"
# gmm-latgen-faster --max-active=7000 --beam=15.0 --lattice-beam=12.0 --acoustic-scale=0.083333 --allow-partial=true --word-symbol-table=exp/monosscf/graph/words.txt exp/monosscf/final.mdl exp/monosscf/graph/HCLG.fst "ark,s,cs:apply-cmvn  --utt2spk=ark:data/test/split2/1/utt2spk scp:data/test/split2/1/cmvn.scp scp:data/test/split2/1/feats.scp ark:- | add-deltas --delta-order=0 ark:- ark:- |" "ark:|gzip -c > exp/monosscf/decode/lat.1.gz"
# Started at Thu Jan  5 13:50:09 ICT 2017
#
gmm-latgen-faster --max-active=7000 --beam=15.0 --lattice-beam=12.0 --acoustic-scale=0.083333 --allow-partial=true --word-symbol-table=exp/monosscf/graph/words.txt exp/monosscf/final.mdl exp/monosscf/graph/HCLG.fst 'ark,s,cs:apply-cmvn  --utt2spk=ark:data/test/split2/1/utt2spk scp:data/test/split2/1/cmvn.scp scp:data/test/split2/1/feats.scp ark:- | add-deltas --delta-order=0 ark:- ark:- |' 'ark:|gzip -c > exp/monosscf/decode/lat.1.gz'
apply-cmvn --utt2spk=ark:data/test/split2/1/utt2spk scp:data/test/split2/1/cmvn.scp scp:data/test/split2/1/feats.scp ark:-

add-deltas --delta-order=0 ark:- ark:-

F05-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F05_Dialog481_MIC01
LOG (gmm-latgen-faster:RebuildRepository():determinize-lattice-pruned.cc:283) Rebuilding repository.
LOG (gmm-latgen-faster:RebuildRepository():determinize-lattice-pruned.cc:283) Rebuilding repository.
WARNING (gmm-latgen-faster:CheckMemoryUsage():determinize-lattice-pruned.cc:316) Did not reach requested beam in determinize-lattice: size exceeds maximum 50000000 bytes; (repo,arcs,elems) = (39511104,23648,11891472), after rebuilding, repo size was 30753088, effective beam was 5.73338 vs. requested beam 12
WARNING (gmm-latgen-faster:DeterminizeLatticePruned():determinize-lattice-pruned.cc:1281) Effective beam 5.73338 was less than beam 12 * cutoff 0.5, pruning raw lattice with new beam 8.29461 and retrying.
LOG (gmm-latgen-faster:RebuildRepository():determinize-lattice-pruned.cc:283) Rebuilding repository.
LOG (gmm-latgen-faster:RebuildRepository():determinize-lattice-pruned.cc:283) Rebuilding repository.
WARNING (gmm-latgen-faster:CheckMemoryUsage():determinize-lattice-pruned.cc:316) Did not reach requested beam in determinize-lattice: size exceeds maximum 50000000 bytes; (repo,arcs,elems) = (32592992,93472,17342712), after rebuilding, repo size was 28398048, effective beam was 8.16923 vs. requested beam 8.29461
WARNING (gmm-latgen-faster:DecodeUtteranceLatticeFaster():decoder-wrappers.cc:273) Determinization finished earlier than the beam for utterance F05-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F05_Dialog481_MIC01
LOG (gmm-latgen-faster:DecodeUtteranceLatticeFaster():decoder-wrappers.cc:285) Log-like per frame for utterance F05-VNSPEECHCORPUSPLUS_ReadSpeech_Studio_F05_Dialog481_MIC01 is -2.66642 over 372 frames.
....
"
and the status in run program is:
"
decode.sh: feature type is delta
bash: line 1: 13184 Killed                  ( gmm-latgen-faster --max-active=7000 --beam=15.0 --lattice-beam=12.0 --acoustic-scale=0.083333 --allow-partial=true --word-symbol-table=exp/monosscf/graph/words.txt exp/monosscf/final.mdl exp/monosscf/graph/HCLG.fst "ark,s,cs:apply-cmvn  --utt2spk=ark:data/test/split2/2/utt2spk scp:data/test/split2/2/cmvn.scp scp:data/test/split2/2/feats.scp ark:- | add-deltas --delta-order=0 ark:- ark:- |" "ark:|gzip -c > exp/monosscf/decode/lat.2.gz" ) 2>> exp/monosscf/decode/log/decode.2.log >> exp/monosscf/decode/log/decode.2.log
run.pl: 2 / 2 failed, log is in exp/monosscf/decode/log/decode.*.log
steps/align_si.sh --nj 2 --cmd run.pl --mem 2G data/train data/lang exp/monosscf exp/monosscf_ali
steps/align_si.sh: feature type is delta
steps/align_si.sh: aligning data in data/train using model from exp/monosscf, putting alignments in exp/monosscf_ali
steps/align_si.sh: done aligning data.

"

So, Could you have any advice for me? Thank you so much

Hang Phuong

[Daniel Povey]

I recommended a wider beam for alignment; you can use a normal beam for decoding.

It was killed because you didn't have enough memory.

[hang phuong Nguyen]

Wow, thank you for you reply quickly. So I will decode with a normal beam.

Thank you so much again

Hang Phuong

[hang phuong Nguyen]

Mr Dan,

Could I ask other question? I trained monophone by my features with 6 dimensions and use beam = 12 and after 40 interactions, I got the content in update.39.log:
"
# gmm-est --write-occs=exp/monosscf/40.occs --mix-up=983 --power=0.25 exp/monosscf/39.mdl "gmm-sum-accs - exp/monosscf/39.*.acc|" exp/monosscf/40.mdl
# Started at Thu Jan  5 06:46:00 ICT 2017
#
gmm-est --write-occs=exp/monosscf/40.occs --mix-up=983 --power=0.25 exp/monosscf/39.mdl 'gmm-sum-accs - exp/monosscf/39.*.acc|' exp/monosscf/40.mdl
gmm-sum-accs - exp/monosscf/39.1.acc exp/monosscf/39.2.acc
LOG (gmm-sum-accs:main():gmm-sum-accs.cc:63) Summed 2 stats, total count 3.21655e+06, avg like/frame -31.3431
LOG (gmm-sum-accs:main():gmm-sum-accs.cc:66) Total count of stats is 3.21655e+06

LOG (gmm-sum-accs:main():gmm-sum-accs.cc:67) Written stats to -

LOG (gmm-est:MleUpdate():transition-model.cc:393) TransitionModel::Update, objf change is 0 per frame over 3.21655e+06 frames.
LOG (gmm-est:MleUpdate():transition-model.cc:396) 110 probabilities floored, 365 out of 577 transition-states skipped due to insuffient data (it is normal to have some skipped.)
LOG (gmm-est:main():gmm-est.cc:102) Transition model update: Overall 0 log-like improvement per frame over 3.21655e+06 frames.

WARNING (gmm-est:MleDiagGmmUpdate():mle-diag-gmm.cc:365) Gaussian has too little data but not removing it because it is the last Gaussian: i = 0, occ = 0, weight = 1
WARNING (gmm-est:MleDiagGmmUpdate():mle-diag-gmm.cc:365) Gaussian has too little data but not removing it because it is the last Gaussian: i = 0, occ = 0, weight = 1

LOG (gmm-est:MleAmDiagGmmUpdate():mle-am-diag-gmm.cc:225) 0 variance elements floored in 0 Gaussians, out of 1025

LOG (gmm-est:MleAmDiagGmmUpdate():mle-am-diag-gmm.cc:229) Removed 0 Gaussians due to counts < --min-gaussian-occupancy=10 and --remove-low-count-gaussians=true

LOG (gmm-est:main():gmm-est.cc:113) GMM update: Overall 0.00912203 objective function improvement per frame over 3.21655e+06 frames
LOG (gmm-est:main():gmm-est.cc:116) GMM update: Overall avg like per frame = -31.3431 over 3.21655e+06 frames.
LOG (gmm-est:SplitByCount():am-diag-gmm.cc:116) Split 143 states with target = 983, power = 0.25, perturb_factor = 0.01 and min_count = 20, split #Gauss from 1025 to 1025
LOG (gmm-est:main():gmm-est.cc:146) Written model to exp/monosscf/40.mdl
# Accounting: time=0 threads=1
# Ended (code 0) at Thu Jan  5 06:46:00 ICT 2017, elapsed time 0 seconds
"

So Were the alignments failing? If yes, Could you have any advice for me?

Thank you again.

Hang Phuong

[Daniel Povey]

it was probably OK but I don't have time to answer more questions.

[hang phuong Nguyen]

Yes, I understand and many thanks for your answers.

Thank you so much.

Hang Phuong