In this paper, we demonstrate the significance of nonlinear neural network models for compression of feature vectors and also develop classifiers for syllable-like units. We consider the standard 80 Stop Consonant-Vowel units of most Indian languages. This set consists of dynamic sounds and hence require large size feature vector to represent the acoustic characteristics of these units. To develop classifiers with limited training data, it is necessary to compress the size of the feature vector. We show that nonlinear compression by autoassociative neural network model is useful, and is superior to the compression by linear principal component analysis.