In this paper, fuzzy information embedded in the MLP classifier is analyzed. Membership function, which represents the degree to which an input pattern belongs to a class, is derived from network output. Analysis show that our membership feature vector can effectively represent the grade of membership without being influenced by the sharp decision boundary formed by MLP trained as classifier. New hybrid system integrating MLP into Hidden Markov Model (HMM) is also proposed 1 and a French vowel recognition experiment has been done based on this hybrid system for comparison of membership feature vector with acoustic feature vector.