In an attempt to make a better model, I changed $CFG_STATESPERHMM in SphinxTrain to 3 from 5. This died in step 05 so I modified that to use 3 states instead of 5 by dropping the last 2 terms of -stwt 1.0 0.3 0.1 0.01 0.001
This seemed to work until I got to step 09 where I received: ges.mk_s2sendump.log:FATAL_ERROR: "mk_s2sendump.c", line 195: #States(3) != 5
Well, now. How am I supposed to use a 3 HMM model if SphinxTrain won't let me create it?
Suggestions on getting around this last problem?
Thanks!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi, I am very interrested in the solution
you may have found. Because I suppose (and maybe
I make a mistake here) that 3 states instead of
5 states will:
-give better recognition on small vocab (500 words)
-make files smaller (sendump, phone, map, etc.), thus making sphinx a good candidate for embedded devices.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Sphinx2 has a fixed 5-state per phone topology and will not run with three states. Sphinx3 will use variable topogies; you can check out s3.0 for that. Sphinx3 code is on sourceforge, though it's not ready for wide release.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
In an attempt to make a better model, I changed $CFG_STATESPERHMM in SphinxTrain to 3 from 5. This died in step 05 so I modified that to use 3 states instead of 5 by dropping the last 2 terms of -stwt 1.0 0.3 0.1 0.01 0.001
This seemed to work until I got to step 09 where I received: ges.mk_s2sendump.log:FATAL_ERROR: "mk_s2sendump.c", line 195: #States(3) != 5
Well, now. How am I supposed to use a 3 HMM model if SphinxTrain won't let me create it?
Suggestions on getting around this last problem?
Thanks!
Hi, I am very interrested in the solution
you may have found. Because I suppose (and maybe
I make a mistake here) that 3 states instead of
5 states will:
-give better recognition on small vocab (500 words)
-make files smaller (sendump, phone, map, etc.), thus making sphinx a good candidate for embedded devices.
Sphinx2 has a fixed 5-state per phone topology and will not run with three states. Sphinx3 will use variable topogies; you can check out s3.0 for that. Sphinx3 code is on sourceforge, though it's not ready for wide release.