In this paper, we constructed a system for realizing a new style of dance performance that dancers play music by dancing. From pilot study, we have found that the motion recognition for dance performance needed the synchronism to back ground music (BGM). Therefore, we propose a new motion recognition method specialized to dance performances. The key techniques of the proposed method are (1) adaptive decision of the size of recognition window to recognize a motion in sync with BGM, and (2) motion recognition in two-phase (rough and detailed) to fulfill the accuracy in high speed recognition. Data was recorded using a 3-axis wireless accelerometers mounted on both shoes. We evaluated the method on a dataset of 5 different dance steps (each repeated 100 times).The results show that this method is capable of improving recognition for all steps (in one case improving recognition from 62% to 99%) while retaining a feeling of seamless connection between movement and sound.