This paper here describes an alternating automaton which can recognize visibly pushdown langauges without using a stack. Unfortunately the transformation from NVPA to such an automaton is skipped in this paper. It also says:
Due to the lack of space, for the omitted details we refer the interested reader to a forthcoming extended version of this paper.
but I actually can not find such a version.
Do you maybe have an alternative paper which explains this in more detail?