The tone variation patterns is very important for speech synthesis. To extract more accurate patterns, a prototype system called SpeechDM has been implemented for learning. In this paper, the relative research and data mining are specified at first, the modules are introduced in detail, and some experimental results are also given.