Classification of taiwanese tones based on pitch and energy movements

This paper addresses the difficulties associated with automatically distinguishing the seven Taiwanese tones. The tone recogniser is an essential component of any automatic speech rec ognition system customised for tone languages such as Taiwanese. We show that it is difficult to distinguish between the Taiwanese tones simply employing the fundamental frequency contours and that the task is simplified by employing energy contour features besides the fundamental frequency features. To allow energy to be accommodated into the classification model an energy-contour feature extraction approach is presented. The proposed approach is inspired by the ADSR model employed in musical instrument synthesis where the envelopes of complex sounds are modeled employing only a few parameters. Our experiments demonstrate that the inclusion of energy into the recognition model allows the seven Taiwanese tones to be discriminated successfully. The paper also presents acoustical measurements of the fundamental frequency and energy features described.