In human locomotion, we continuously modulate joint mechanical impedance of the lower limb (hip, knee, and ankle) either voluntarily or reflexively to accommodate environmental changes and maintain stable interaction. Ankle mechanical impedance plays a pivotal role at the interface between the neuro-mechanical system and the physical world. This paper reports, for the first time, a characterization of human ankle mechanical impedance in two degrees-of-freedom simultaneously as it varies with time during walking. Ensemble-based linear time-varying system identification methods implemented with a wearable ankle robot, Anklebot, enabled reliable estimation of ankle mechanical impedance from the pre-swing phase through the entire swing phase to the early-stance phase. This included heel-strike and toe-off, key events in the transition from the swing to stance phase or vice versa. Time-varying ankle mechanical impedance was accurately approximated by a second order model consisting of inertia, viscosity, and stiffness in both inversion-eversion and dorsiflexion-plantarflexion directions, as observed in our previous steady-state dynamic studies. We found that viscosity and stiffness of the ankle significantly decreased at the end of the stance phase before toe-off, remained relatively constant across the swing phase, and increased around heel-strike. Closer investigation around heel-strike revealed that viscosity and stiffness in both planes increased before heel-strike occurred. This finding is important evidence of "pretuning" by the central nervous system. In addition, viscosity and stiffness were greater in the sagittal plane than in the frontal plane across all subgait phases, except the early stance phase. Comparison with previous studies and implications for clinical study of neurologically impaired patients are provided.