This paper describes the use of a lexical tree based lookahead for reducing the computation needed by the lexical search stage of a segment-based speech recognition system. In the MIT SUMMIT system, a network representing possible phonetic interpretations of the signal is generated before the lexical search is performed. This allows the use of a fairly simple tree-based lookahead in order to find a reduced set of words which may be allowed to start at any point in time. The first N phones in the pronunciations of all words in the lexicon are collapsed into a tree and this tree is matched against the segment network using a pruning threshold to determine a subset of words which may start at each node in the segment network. We describe the computational needs of the system, explains the lookahead in more detail, and shows the tradeoffs between computation and accuracy.