A key question in automatic assessment is how pronunciation can be measured to determine its appropriateness. One method is to use the spectral features of compactness, graveness, stress, and flatness to assess specific sounds. The decision is made based on a supervisor's perceptual analysis from formants and energy in spectral bands. Mathematical rule can also be used to measure a dissimilarity or distance between the closeness of the input feature vector and a set of predefined feature vectors[18].
