This paper describes Romaine, a lexical access model designed to address the potential shortcomings of some lexical access approaches, namely, flat lexical representation and left-to-right search. The key ideas embodied in Romaine are a compact hierarchical representation of the lexicon, and a bottom-up, island-driven control strategy that clusters constituent hypotheses into aggregate sequences. We also discuss the results of a set of simulation experiments designed to help us understand the performance of the system as vocabulary size grows from 200 to 20,000 words. These results suggests that Romaine is computationally very efficient, and that the accuracy remains reasonable when segmentation and classifications errors are introduced.