I needed some help regarding training of continuous models for sphinxIII. I am trying to use the sphinx trainer package to train phonetic models. Currently, I am trying this with very small training data set(4-5 words), and the required phones for this data set. I had carried out the steps given in the training document(for continuous models), and accordingly obtained the resulting files. I then tried to use these files with SphinxIII decoder. But the decoder gives access violation error and dies. The obtained output is given below. We are using the decoder given for windows with the sample batch file "sphinx3-test.bat", with the relevant changes done in the "args.an4.test" file. I have replaced the required files in the folders. Also I generated a language model file (using the tool available on the website) which includes a subset of words taken during training. Then created a DMP file using lm3g2dmp (windows version in SphinxII package). The utterance file to be decoded is also the same one as that taken during training.The training process generated all the necessary files with some warnings, but without any errors.
Note: The trainer is run in Linux and we are trying to decode in windows using the files generated in Linux.
Thanks in advance,
Parag
OBTAINED OUTPUT:-
C:\Sphinx_test\sphinx3-0.1\win32\batch>echo off
.
sphinx3-test
Run CMU Sphinx-3 in Batch mode to decode an example utterance.
.
This batch script assumes all files are relative to the main
directory (S3ROOT).
.
When running this, look for a line that starts with "FWDVIT:"
If the installation is correct, this line should read:
FWDVID: P I T T S B U R G H (null)
INFO: c:\sphinx_test\sphinx3-0.1\src\libutil\cmd_ln.c(276): Parsing command line:
\
-mdef ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/hub4opensrc.6000.mdef \
-fdict ./model/lm/an4/filler.dict \
-dict ./model/lm/an4/an4.dict \
-mean ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/means \
-var ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/variances \
-mixw ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/mixture_weights \
-tmat ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/transition_matrices \
-upperf 6855.49756 \
-lowerf 133.33334 \
-nfilt 40 \
-samprate 16000 \
-agc none \
-varnorm no \
-cmn current \
-subvqbeam 1e-02 \
-epl 4 \
-fillprob 0.02 \
-lw 9.5 \
-maxwpf 10 \
-beam 1e-90 \
-wbeam 1e-40 \
-lm ./model/lm/an4/an4.ug.lm.DMP
Hi,
I needed some help regarding training of continuous models for sphinxIII. I am trying to use the sphinx trainer package to train phonetic models. Currently, I am trying this with very small training data set(4-5 words), and the required phones for this data set. I had carried out the steps given in the training document(for continuous models), and accordingly obtained the resulting files. I then tried to use these files with SphinxIII decoder. But the decoder gives access violation error and dies. The obtained output is given below. We are using the decoder given for windows with the sample batch file "sphinx3-test.bat", with the relevant changes done in the "args.an4.test" file. I have replaced the required files in the folders. Also I generated a language model file (using the tool available on the website) which includes a subset of words taken during training. Then created a DMP file using lm3g2dmp (windows version in SphinxII package). The utterance file to be decoded is also the same one as that taken during training.The training process generated all the necessary files with some warnings, but without any errors.
Note: The trainer is run in Linux and we are trying to decode in windows using the files generated in Linux.
Thanks in advance,
Parag
OBTAINED OUTPUT:-
C:\Sphinx_test\sphinx3-0.1\win32\batch>echo off
.
sphinx3-test
Run CMU Sphinx-3 in Batch mode to decode an example utterance.
.
This batch script assumes all files are relative to the main
directory (S3ROOT).
.
When running this, look for a line that starts with "FWDVIT:"
If the installation is correct, this line should read:
FWDVID: P I T T S B U R G H (null)
INFO: c:\sphinx_test\sphinx3-0.1\src\libutil\cmd_ln.c(276): Parsing command line:
\ -mdef ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/hub4opensrc.6000.mdef \ -fdict ./model/lm/an4/filler.dict \ -dict ./model/lm/an4/an4.dict \ -mean ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/means \ -var ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/variances \ -mixw ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/mixture_weights \ -tmat ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/transition_matrices \ -upperf 6855.49756 \ -lowerf 133.33334 \ -nfilt 40 \ -samprate 16000 \ -agc none \ -varnorm no \ -cmn current \ -subvqbeam 1e-02 \ -epl 4 \ -fillprob 0.02 \ -lw 9.5 \ -maxwpf 10 \ -beam 1e-90 \ -wbeam 1e-40 \ -lm ./model/lm/an4/an4.ug.lm.DMP
Configuration in effect:
[NAME] [DEFLT] [VALUE]
-agc max none
-beam 1.0e-55 1.000000e-090
-bghist 0 0
-bptbldir
-cepdir
-cmn current current
-ctl
-ctlcount 1000000000 1000000000
-ctloffset 0 0
-dict ./model/lm/an4/an4.dict
-epl 3 4
-fdict ./model/lm/an4/filler.dict
-fillpen
-fillprob 0.1 2.000000e-002
-hmmdump 0 0
-hmmhistbinsize 5000 5000
-hypseg
-latext lat.gz lat.gz
-lextreedump 0 0
-lm ./model/lm/an4/an4.ug.lm.DMP
-lminmemory 0 0
-log3table 1 1
-logbase 1.0003 1.000300e+000
-lowerf 200 1.333333e+002
-lw 8.5 9.500000e+000
-maxcepvecs 256 256
-maxhistpf 100 100
-maxhmmpf 20000 20000
-maxhyplen 1000 1000
-maxwpf 20 10
-mdef ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/hub4opensrc.6000.mdef
-mean ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/means
-mixw ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/mixture_weights
-mixwfloor 0.0000001 1.000000e-007
-nfilt 31 40
-Nlextree 3 3
-outlatdir
-outlatoldfmt 1 1
-pbeam 1.0e-50 1.000000e-050
-ptranskip 0 0
-samprate 8000 16000
-silprob 0.1 1.000000e-001
-subvq
-subvqbeam 3.0e-3 1.000000e-002
-tmat ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/transition_matrices
-tmatfloor 0.0001 1.000000e-004
-treeugprob 1 1
-upperf 3500 6.855498e+003
-utt
-uw 0.7 7.000000e-001
-var ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/variances
-varfloor 0.0001 1.000000e-004
-varnorm no no
-vqeval 3 3
-wbeam 1.0e-35 1.000000e-040
-wip 0.7 7.000000e-001
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\kbcore.c(90): Initializing core models:
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\logs3.c(99): Initializing logbase: 1.000300e+000 (add table: 1)
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\logs3.c(161): Log-Add table size = 29350
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\feat.c(643): Initializing feature stream to type: '1s_c_d_dd', CMN='current', VARNORM='no', AGC='none'
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\mdef.c(594): Reading model definition: ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/hub4opensrc.6000.mdef
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\mdef.c(771): 14 CI-phone, 14 CD-phone, 3 emitstate/phone, 42 CI-sen, 84 Sen, 28 Sen-Seq
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\dict.c(358): Reading main dictionary: ./model/lm/an4/an4.dict
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\dict.c(361): 3 words read
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\dict.c(366): Reading filler dictionary: ./model/lm/an4/filler.dict
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\dict.c(369): 3 words read
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\lm.c(517): LM read('./model/lm/an4/an4.ug.lm.DMP', lw= 9.50, wip= -1188, uw= 0.70)
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\lm.c(333): 5 unigrams
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\lm.c(365): 6 bigrams [on disk]
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\lm.c(373): 3 trigrams [on disk]
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\lm.c(395): 3 bigram prob entries
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\lm.c(413): 3 trigram bowt entries
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\lm.c(429): 2 trigram prob entries
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\lm.c(444): 1 trigram segtable entries (512 segsize)
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\lm.c(478): 5 word strings
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(88): Reading mixture gaussian file './model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/means'
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(212): 84 mixture Gaussians, 1 components, veclen 39
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(88): Reading mixture gaussian file './model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/variances'
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(212): 84 mixture Gaussians, 1 components, veclen 39
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(233): Reading mixture weights file './model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/mixture_weights'
ERROR: "c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c", line 306: Weight normalization failed for 3 senones
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(318): Read 84 x 1 mixture weights
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(334): Removing uninitialized Gaussian densities
11 24 27 28 29 53 64 69
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(364): 8 densities removed (8 mixtures removed entirely)
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(372): Applying variance floor
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(384): 103 variance values floored
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cont_mgau.c(428): Precomputing Mahalanobis distance invariants
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\tmat.c(135): Reading HMM transition probability matrices: ./model/hmm/hub4_cd_continuous_8gau_1s_c_d_dd/transition_matrices
ERROR: "c:\sphinx_test\sphinx3-0.1\src\libs3decoder\tmat.c", line 197: Normalization failed for tmat 9 from state 0
ERROR: "c:\sphinx_test\sphinx3-0.1\src\libs3decoder\tmat.c", line 197: Normalization failed for tmat 9 from state 1
ERROR: "c:\sphinx_test\sphinx3-0.1\src\libs3decoder\tmat.c", line 197: Normalization failed for tmat 9 from state 2
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\tmat.c(217): Read 14 transition matrices of size 3x4
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\dict2pid.c(254): Building PID tables for dictionary
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\dict2pid.c(422): 21 composite states; 7 composite sseq
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\kbcore.c(172): Verifying models consistency:
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\kb.c(141): Building lextrees
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\kb.c(207): Lextrees(3), 10 nodes(ug), 1 nodes(filler)
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\kb.c(229): Beam= -690764, PBeam= -383758, WBeam= -307006, SVQBeam= -15350
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\vithist.c(71): Initializing Viterbi-history module
45056 samples in file .\model\lm\an4/pittsburgh.littleendian.raw.
Will be decoded in blocks of 2000
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\feat.c(943): Feature buffers initialized to 256 vectors
INFO: c:\sphinx_test\sphinx3-0.1\src\libs3decoder\cmn_prior.c(72): mean[0]= 12.00, mean[1..12]= 0.0
INFO: c:\sphinx_test\sphinx3-0.1\src\programs\live.c(239): live_nfeatvec: 7
ERROR: "c:\sphinx_test\sphinx3-0.1\src\libs3decoder\utt.c", line 655: ***ERROR*** Fr 0, best HMM score > 0 (1476740842); int32 wraparound?
ERROR: "c:\sphinx_test\sphinx3-0.1\src\libs3decoder\utt.c", line 655: ***ERROR*** Fr 1, best HMM score > 0 (537216746); int32 wraparound?