J. Takahashi et S. Sagayama, DISCRIMINATIVE TRAINING BASED ON MINIMUM CLASSIFICATION ERROR FOR A SMALL AMOUNT OF DATA ENHANCED BY VECTOR-FIELD-SMOOTHED BAYESIAN LEARNING, IEICE transactions on information and systems, E79D(12), 1996, pp. 1700-1707
This paper describes how to effectively use discriminative training ba
sed on Minimum Classification Error (ME) criterion for a small amount
of data in order to attain the highest Level of recognition performanc
e. This method is a combination of MCE training and Vector-Field-Smoot
hed Bayesian learning called MAP/VFS, which combines maximum a posteri
ori (MAP) estimation with Vector Field Smoothing (VFS). In the propose
d method, MAP/VFS can significantly enhance MCF training in the robust
ness of acoustic modeling. In model training, MCE training is performe
d using the MAP/VFS-trained model as an initial model. The same data a
re used in both trainings. For speaker adaptation using several dozen
training words, the proposed method has been experimentally proven to
be very effective. For 50-word training data, recognition errors are d
rastically reduced by 47% compared with 16.5% when using only MCE. Thi
s high rate, in which 39% is due to MAP, an additional 4% is due to VF
S, and a further improvement of 4% is due to MCE, can be attained by e
nhancing MCE training capability by MAP/VFS.