Subsyllabic Tone Units for Reducing Physiological Effects in Automatic Tone Recognition for Connected Mandarin

This paper presents our attempt to model physiological transition effect on syllable F0 contour in order to improve lexical tone recognition performance for Mandarin Chinese. We suggested that a syllable F0 contour consists of three segments: onset course, tone nucleus and offset course. Among the three segments, only tone nucleus contains key features for tone recognition, and the other two result from physiological transition effect of human vocal cords. Therefore, the tone recognizer performance can be improved by only focusing on tone nuclei and discarding the other two segments. Segmentation of the three segments was achieved by our proposed method. Contextdependent tonal models, which are trained on tone nucleus features, were also introduced to model contextual tone coarticulation effects for tone recognition. Advantages of the proposed methods were proved through tone recognition experiments of continuous speech of Mandarin.