Gesture, especially finger gestures, can reflect various intentions of users in daily behaviors. Therefore, more and more researches are performed on recognition of finger gestures based on wearable sensors. Segmentation, detecting the Start and End of a gesture, is crucial for accurate recognition of the gesture. However it is hard to extract signal of each finger gesture timely and correctly from continuous signals of the sensor, since they are transient, noise sensitive, and with individual difference. In this paper, we design an adaptive separator employing Bayes theory and distribution statistic technology to overcome the difficulties. The separator can detect each segmentation from a series of gestures timely, correctly, and adaptive to each user.