We describe a new approach for rescoring speech lattices - with long-span language models or wide-context acoustic models - that does not entail computationally intensive lattice expansion or limited rescoring of only an N-best list. We view the set …