Abstract : This paper deals with two new algorithms for Multi-Band Speech Recognition : the first one enhances the robustness of the system to stationary noise, and the second one is a global training algorithm which allows the system to reach its global optimum. Experimental results confirm the good behavior of these two methods.