Exploring Tone Variations in Chinese Dialects Using Context Dependent Tone Models

In this paper we propose a statistical approach for prosody study. It has two key stages: first, context dependent models are trained automatically from a natural speech corpus. Then data mining and data visualization techniques are used to discover phonetic knowledge from the model parameters. We use this approach to study the tonal system of Xi´an dialect, where we learn much knowledge about the dialect, without the need of any specific phonetic annotations about it. Some of our observations coincide with those from other studies, which demonstrate the capability of this approach, while some of our new findings show its advantages comparing with traditional methods.