A comparative study of a scalar waveform modelling approaches based on linear, quadratic and neural prediction techniques was performed. A linear spectral predictor model was also studied on two spectral representations. For this, we consider the task of discriminating the places of articulation of stop consonants using their spectral transition information. A modified error measure has been defined and shown to be effective in reducing the pitch interference effect that arise in directly modelling the waveform. The importance of complete spectral information of the vocal tract system dynamics and the significance of higher order correlation in the recognition performance have been highlighted.