hi. i have about 200 utterances for adaptation from one speaker. i divided them into two sets (Set A and B)
-> i adapted(http://www.cs.cmu.edu/~archan/presentation/MAP.pdf) the generic acoustic model with set A and then adapted the updated model with set B. (lets call the updated model as 'M1')
-> i combined Sets A and B and then adapted the generic acoustic model with the combined utterances. (lets call the adapted model as 'M2')
i have utterances from the same speaker that are not included in Set A and Set B and used them to test M1 and M2.
i noticed that M2 gives better accuracy than M1. why is this so? when both m1 and m2 are adapted from the same data.
thank you
-kris
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
so, if i have another set of utterances, set C, it would be better if i combined set C with sets A and B and then adapt them to the generic acoustic model, than to adapt set C to either M1 or M2. is this correct?
thanks.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Not very strange. M1 is overtrained to B train set which is badly generalize to the test set. M2 is trained to A+B and they are closer to the test set.
Remember that training is an estimation of distribution. The more data you have the better the accuracy. The same is true for adaptation.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
hi. i have about 200 utterances for adaptation from one speaker. i divided them into two sets (Set A and B)
-> i adapted(http://www.cs.cmu.edu/~archan/presentation/MAP.pdf) the generic acoustic model with set A and then adapted the updated model with set B. (lets call the updated model as 'M1')
-> i combined Sets A and B and then adapted the generic acoustic model with the combined utterances. (lets call the adapted model as 'M2')
i have utterances from the same speaker that are not included in Set A and Set B and used them to test M1 and M2.
i noticed that M2 gives better accuracy than M1. why is this so? when both m1 and m2 are adapted from the same data.
thank you
-kris
thanks nickolay..
so, if i have another set of utterances, set C, it would be better if i combined set C with sets A and B and then adapt them to the generic acoustic model, than to adapt set C to either M1 or M2. is this correct?
thanks.
Looks so
Not very strange. M1 is overtrained to B train set which is badly generalize to the test set. M2 is trained to A+B and they are closer to the test set.
Remember that training is an estimation of distribution. The more data you have the better the accuracy. The same is true for adaptation.