Hello friends.
I´m trying -tau parameter on adaptation, (ALPHA 5 VER)
Does it work ?
I´m using big french model adapted for me (ACCURACY 78%, for standard 167 unadapted sentences)
I see on forum to test param >100 ??!!
What value to start ? 120% ?
How to test it impact ? accuracy ? speed ?
Could anyone help me
thanks guys
Vincent
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
THANKS NICKOLAY
So -bayesmean no
-tau tested with different values (5, 30, 90,120,160)
When -fixedtau no, i always have same accuracy after adapt
with -fixedtau yes, accuracy value change
But I never have better than with -BAYESMEAN YES (standard map_adapt, without variances)
Is it normal ?
I was thinking that -tau usage could improve accuracy. not the case
Appreciate your help friends
vincent
Last edit: Elpimous 2016-05-09
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Well, you probably won't get too much from simple MAP. If you have enough data it's probably better to retrain or even train new DNN acoustic models with Kaldi. French voxforge is quite big already actually to train a new model.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello friends.
I´m trying -tau parameter on adaptation, (ALPHA 5 VER)
Does it work ?
I´m using big french model adapted for me (ACCURACY 78%, for standard 167 unadapted sentences)
I see on forum to test param >100 ??!!
What value to start ? 120% ?
How to test it impact ? accuracy ? speed ?
Could anyone help me
thanks guys
Vincent
You can test adaptation on a test set as described in testing part of the tutorial
http://cmusphinx.sourceforge.net/wiki/tutorialadapt#testing_the_adaptation
You can try values from 10 to 200
I tried values 60, 100, 130 and 160.
No difference on accuracy
anything like if tau parameter didn´t work
can anyone confirm ?
Tau works only if bayesmean is no.
THANKS NICKOLAY
So -bayesmean no
-tau tested with different values (5, 30, 90,120,160)
When -fixedtau no, i always have same accuracy after adapt
with -fixedtau yes, accuracy value change
But I never have better than with -BAYESMEAN YES (standard map_adapt, without variances)
Is it normal ?
I was thinking that -tau usage could improve accuracy. not the case
Appreciate your help friends
vincent
Last edit: Elpimous 2016-05-09
hi friends,
anyone could give me correct direction ?
For now, i have very poor results with -tau
Hi
Well, you probably won't get too much from simple MAP. If you have enough data it's probably better to retrain or even train new DNN acoustic models with Kaldi. French voxforge is quite big already actually to train a new model.