Abstract: The lack of training samples is an imperative problem in the research of sign language recognition. In this paper, a method of using derivative data is proposed, which facilitates synthesizing dynamic samples of multi-stream employed in sign language training. With the mean-shift algorithm, the movement direction of density function grads can be obtained easily, thus controling the direction and the intensity of derivation. At the same time, this algorithm can also satisfy the need for the synthesized samples to include as much and as effective information of unspecific signer as possible. Moreover, the method realizes a transformation of data which is irrevocable in the initialization process of the recognition system. The driving effect of the synthesized data depends on the capacity of the model, as well as the intensity and direction of synthesization. After assessing the driving effect under various experiment environments, it is found that in most cases the recognition rate is raised; and in some cases, it is even markedly raised.