hi
i have few question about MLLR adaptation using sphinx train for telephonic speech ,i built models from all the corpus ,next step was to try to adapt the models to the telephonic speech base.
first i done a single itération of baum welch with the models using adaptation data .but here , which argument of the executable ./bw must i put ?
can i use the same arguments , and let -varnorm to yes......, or retrain all the baseline with adaptation arguments .?
can some one show me the exact methode for mllr adaptation startin to the begining ( model from all train data )?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I used the command as found in the link Arthur provided but with the same accumelator directory used for training the old dat, it gave me no errors but produced an idintity matrix. So I decided to use all the arguments found in the file 04.cd_schmm_untied/baum_welch.pl, it didn't work either giving me the same results ????! How do I know my adaptation was successful
hi
i tried the command as found in the link (given by arthur) but always i have lot of errors :
ERROR: "backward.c", line 409: final state not reached
ERROR: "baum_welch.c", line 300: 19981207_0700_0800/19981207_0700_0800.251 ignored
about 75 % of data are ignored , i don't know why ?
i saw in the forum but i didn't found an exact explanation .
please if some one uderstand what happened !!
thanks
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
hi
i have few question about MLLR adaptation using sphinx train for telephonic speech ,i built models from all the corpus ,next step was to try to adapt the models to the telephonic speech base.
first i done a single itération of baum welch with the models using adaptation data .but here , which argument of the executable ./bw must i put ?
can i use the same arguments , and let -varnorm to yes......, or retrain all the baseline with adaptation arguments .?
can some one show me the exact methode for mllr adaptation startin to the begining ( model from all train data )?
http://www.cs.cmu.edu/~archan/sphinxDoc.html
-a
I used the command as found in the link Arthur provided but with the same accumelator directory used for training the old dat, it gave me no errors but produced an idintity matrix. So I decided to use all the arguments found in the file 04.cd_schmm_untied/baum_welch.pl, it didn't work either giving me the same results ????! How do I know my adaptation was successful
the command:
bin/bw -moddeffn \"$moddeffn\" -mixwfn \"$mixwfn\" -mwfloor $mwfloor -ltsoov $CFG_LTSOOV -2passvar yes -topn 5 -agc $CFG_AGC -cmn $CFG_CMN -varreest yes -tmatreest yes -ceplen $CFG_VECTOR_LENGTH -varfloor $minvar -meanfn \"$meanfn\" -varfn \"$varfn\" -tmatfn \"$tmatfn\" -dictfn \"$CFG_DICTIONARY\" -fdictfn \"$CFG_FILLERDICT\" -cepdir \"$CFG_FEATFILES_DIR\" -ts2cbfn \"$statepdeffn\" -cepext $CFG_FEATFILE_EXTENSION -lsnfn \"$transcriptfile\" -meanreest yes -feat 1s_c_d_dd -ceplen $CFG_VECTOR_LENGTH -ctlfn \"$ADBT_CTL\" -accumdir \"$output_buffer_dir\" -abeam 1e-90 -bbeam 1e-40
hi
i tried the command as found in the link (given by arthur) but always i have lot of errors :
ERROR: "backward.c", line 409: final state not reached
ERROR: "baum_welch.c", line 300: 19981207_0700_0800/19981207_0700_0800.251 ignored
about 75 % of data are ignored , i don't know why ?
i saw in the forum but i didn't found an exact explanation .
please if some one uderstand what happened !!
thanks