I'm doing isolated word recognition, hence training CI models only. For cross-
validation, I need to retrain word models over and over again, using slightly
modified training sets. During that process, only few models change, all
others remain unchanged. However, it seems to me that the SphinxTrain setup
doesn't allow me to retrain a selection of CI models only, unlike HTK for
instance, where the CI word models are stored and trained separately, which
allows retraining only those with a modified training set. Isn't there a way
to tell SphinxTrain "hey, by the way, only HMM no. XX needs retraining, use
what you already have for all others"?
Thank you for any hints!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hm, that's a pity. Do you think it would be feasible to train each word
separately (with an ultra-short dictionary etc. consisting of only that word)
and then (by own script) combine the means binary files, variances binary
files etc. into joint full-size files containg all my word models, as if they
had been trained together? I would probably need a file format reference
document...
Right now, I'm just having 99.9% unnecessary redundant retraining overhead...
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi
I'm doing isolated word recognition, hence training CI models only. For cross-
validation, I need to retrain word models over and over again, using slightly
modified training sets. During that process, only few models change, all
others remain unchanged. However, it seems to me that the SphinxTrain setup
doesn't allow me to retrain a selection of CI models only, unlike HTK for
instance, where the CI word models are stored and trained separately, which
allows retraining only those with a modified training set. Isn't there a way
to tell SphinxTrain "hey, by the way, only HMM no. XX needs retraining, use
what you already have for all others"?
Thank you for any hints!
No, there is no such option.
Hm, that's a pity. Do you think it would be feasible to train each word
separately (with an ultra-short dictionary etc. consisting of only that word)
and then (by own script) combine the means binary files, variances binary
files etc. into joint full-size files containg all my word models, as if they
had been trained together? I would probably need a file format reference
document...
Right now, I'm just having 99.9% unnecessary redundant retraining overhead...