3.2 Estimation of Confidence Scores for the Word Hypotheses
The verification scheme is based on the estimation of character probabilities by a multilayer perceptron (MLP) neural network which is called segmental neural network
(SNN). The architecture of the SNN resembles a standard MLP character classifier. The task of the SNN is to assign an a posteriori probability to each segment representing a
character given that the character class has already been assigned by the HRS. We define xl as the feature vector corresponding to the lth word segment and cl as the character
class of the lth word segment provided by the HRS. The output of the SNN is PðcljxlÞ, which is the a posteriori probability of the character class cl given the feature vector xl.