-lw 15
-feat 1s_c_d_dd
-wip 0.2
-beam 1e-120
-pbeam 1e-120
-wbeam 1e-100
-varnorm no
-cmn current
but still no words were recognized. Replacing wsj's mdef with http://www.cs.cmu.edu/~dhuggins/Projects/wsj_all_cd30.mllt.4000.mdef made no difference. Using "-fdict /home/bshanks/prog/speech/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.filler" made no difference. Leaving out "-lda /home/bshanks/prog/speech/wsj_all_cd30.mllt_cd_cont_4000/feature_transform" made no difference.
Hi,
I followed http://sphinx.subwiki.com/sphinx/index.php/Hello_World_Decoder_QuickStart_Guide successfully and am now trying to use wsj rather than hub. I am executing "sphinx3_livedecode _CFG" and trying different options in _CFG. When _CFG contains:
-samprate 16000
-dict /home/bshanks/prog/speech/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.dic
-fdict /home/bshanks/prog/speech/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.filler
-lm ./dmp/lm_giga_5k_nvp_3gram.arpa.DMP
-hmm /home/bshanks/prog/speech/hub4_cd_continuous_8gau_1s_c_d_dd
then sphinx3_livedecode recognizes each word in "yellow or red" properly (everything works). But when _CFG contains:
-samprate 16000
-dict /home/bshanks/prog/speech/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.dic
-lm ./dmp/lm_giga_5k_nvp_3gram.arpa.DMP
-hmm /home/bshanks/prog/speech/wsj_all_cd30.mllt_cd_cont_4000
-lda /home/bshanks/prog/speech/wsj_all_cd30.mllt_cd_cont_4000/feature_transform
-fdict wsj_all_cd30.mllt_cd_cont_4000/noisedict
then sphinx3_livedecode does not recognize a single word (sometimes it incorrectly reports "the", but usually it reports nothing).
After reading http://sourceforge.net/forum/forum.php?thread_id=2000591&forum_id=5471 , I tried:
-samprate 16000
-dict /home/bshanks/prog/speech/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.dic
-lm ./dmp/lm_giga_5k_nvp_3gram.arpa.DMP
-hmm /home/bshanks/prog/speech/wsj_all_cd30.mllt_cd_cont_4000
-lda /home/bshanks/prog/speech/wsj_all_cd30.mllt_cd_cont_4000/feature_transform
-fdict wsj_all_cd30.mllt_cd_cont_4000/noisedict
-lw 15
-feat 1s_c_d_dd
-wip 0.2
-beam 1e-120
-pbeam 1e-120
-wbeam 1e-100
-varnorm no
-cmn current
but still no words were recognized. Replacing wsj's mdef with http://www.cs.cmu.edu/~dhuggins/Projects/wsj_all_cd30.mllt.4000.mdef made no difference. Using "-fdict /home/bshanks/prog/speech/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.filler" made no difference. Leaving out "-lda /home/bshanks/prog/speech/wsj_all_cd30.mllt_cd_cont_4000/feature_transform" made no difference.
I would have liked to try an old version of wsj as was done in http://sourceforge.net/forum/forum.php?thread_id=2000591&forum_id=5471 , but I do not know where to get one -- http://www.speech.cs.cmu.edu/sphinx/models only links to the new version.
I am using trunk, and
http://www.speech.cs.cmu.edu/sphinx/models/wsj_jan2008/wsj_all_mllt_4000_20080104.tar.gz
http://www.speech.cs.cmu.edu/sphinx/models/hub4opensrc_jan2002/4000senones/1s_c_d_dd/hub4_cd_continuous_8gau_1s_c_d_dd.tar.gz
http://www.inference.phy.cam.ac.uk/kv227/lm_giga/lm_giga_5k_nvp_3gram.zip
Did I do something wrong, or is this a problem with the new wsj? Thanks,
bayle
Well, new wsj indeed has problems. I hope it will be updated soon or the code will be fixed. You could try wsj from
http://www.inference.phy.cam.ac.uk/kv227/sphinx/acoustic_models.html
Oh, sorry, for later readers: that .dic and .lm are my own custom. You can also use
-dict lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.dic
-lm ./dmp/lm_giga_5k_nvp_3gram.arpa.DMP
thanks! it works now.
for later readers, i got wsj_all_cont_3no_8000_16 off of http://www.inference.phy.cam.ac.uk/kv227/sphinx/acoustic_models.html and used the following configuration arguments:
-hmm /home/bshanks/prog/speech/wsj_all_cont_3no_8000_16
-dict voice-keyboard-current.dic
-lm voice-keyboard-current.lm.DMP
-fdict /home/bshanks/prog/speech/lm_giga_5k_nvp_3gram/lm_giga_5k_nvp.sphinx.filler
-feat s3_1x39