Use of effective features for emotion recognition is a step towards better accuracy. The segmental features extracted from the utterances have a high dimension as well as redundant. Low-dimensional feature set for neural network based model is essential and equally important. However the reduction technique is to be chosen properly. This paper proposes a feature reduction mechanism using the combination of Vector Quantization (VQ) and eigenvalue decomposition for effective feature utility. The lower order Eigen components are more informative as compared to the Principal Components and are considered in this work to analyze children speech emotions. For classification, the Probabilistic Neural Network (PNN) model is considered due to its accuracy level with the statistical features. It has been observed that the accuracy level is higher with the proposed VQ based Eigen features. Enhanced accuracy up to 97.5% for VQ based Mel-frequency Cepstral coefficients reduced by their eigenvalues has been exhibited.