Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss
UBM Training
59 views
Skip to first unread message
Nicolas Lubimov
Nov 18, 2011, 4:52:08 AM11/18/11
to
Hello!
I have one technical question about construction of Universal
Background Model (UBM). I'm trying to build the speaker recognition
system that uses UBM, and in several articles on the same topic I've
met the statement that UBM needs for very big training size (about
100h of speech), so it is not so trivial to train this model - it is
time and memory consuming operation. I have parallelized my procedure,
it has became a bit faster, but not so satisfactory. May be it is not
the right way of trying to use the whole dataset (in my case it is
about 30 million MFCC points)? I guess that the people have another
way to get UBM faster, for example, by randomly selecting small
subset... But I'm afraid that this method could affect on the quality
of final model, isn't it?
I have one technical question about construction of Universal
Background Model (UBM). I'm trying to build the speaker recognition
system that uses UBM, and in several articles on the same topic I've
met the statement that UBM needs for very big training size (about
100h of speech), so it is not so trivial to train this model - it is
time and memory consuming operation. I have parallelized my procedure,
it has became a bit faster, but not so satisfactory. May be it is not
the right way of trying to use the whole dataset (in my case it is
about 30 million MFCC points)? I guess that the people have another
way to get UBM faster, for example, by randomly selecting small
subset... But I'm afraid that this method could affect on the quality
of final model, isn't it?
Bill Xia
Oct 12, 2012, 10:56:04 AM10/12/12
to
在 2011年11月18日星期五UTC+8下午5时52分08秒，Nikolay Lyubimov写道：
It do affect the quality of final model, but the affection is small. Actually if we randomly choose just 1% of the dataset to train UBM, the final EER only decrease by less than 0.01%. You can make a compromise between the time and final results.
0 new messages