Humans utilize the extended-knee configuration to achieve efficient walk. A humanoid robot, however, cannot do so, since this posture is a singular one, and hence, an extremely large joint velocity would be generated, provided a conventional control method is used. Some other methods exist, though, that are able to handle motion control at such postures. One example is the Singularity Consistent (SC) Approach, which we used in a previous work to generate a static walk for a humanoid robot. In this paper, we propose a new method to generate a dynamic walking pattern through the singularity neighborhood, by using proper ankle control and pattern generation based on a spherical inverted pendulum model. Experimental results show the effectiveness of this approach.