In this paper, we examined different methods of modeling prosodic features of tones, and their effects on a speaker-independent Taiwanese tone recognition system. Tones can be modeled either by plain or curve-fitted features. Plain features represent the original curve faithfully using pitch values, while curve- fitted features can be thought of as an approximation to the values using mathematical functions, such as a Legendre polynomial. In addition, durational information of tones was also proven effective in previous researches. Thus, we proposed a new approach of modeling Taiwanese tones using curve-fitted features extracted from fractions of the pitch curve, along with duration as an additional prosodic feature. Our experimental results showed that using these features in an SVM classifier could substantially improve the accuracy of tone recognition in Taiwanese. Besides, we provided an empirical perspective for theoretic studies on tonal neutralization.