A simple tone classifier for Cantonese recognition

Abstract

This paper presents a method for tone classification of discrete Cantonese utterance. An in-depth study on the characteristics, properties and variations of different isolated words with the same tone and spoken by various speakers has been carried out. Based on this study, several feature parameters that are pertinent to tone identification were defined. Although some of these parameters are dependent on the intrinsic pitch of the speaker, speaker dependence of the feature set is effectively removed by using a novel intrinsic pitch estimation procedure. A neural net was trained to perform tone classification by utilizing the feature set derived. An average speaker-independent recognition accuracy of around 84% was obtained.<>