Accurate processing of speech requires that listeners map temporally unfolding input to words. A long-held set of principles describes this process: lexical items are activated immediately and incrementally as speech arrives, perceptual and lexical representations rapidly decay to make room for new information; and lexical entries are temporally structured. In this framework; speech processing is tightly coupled to the temporally unfolding input. However, recent work challenges this: low-level auditory and higher-level lexical representations do not decay and are instead retained over long durations, speech perception may require encapsulated memory buffers, lexical representations are not strictly temporally structured, and listeners can substantially delay lexical access in some circumstances. These findings suggest that current theories and models of word recognition need to be reconceptualized.