Entropy is proposed as a robust statistical measure of the performance of the lexical access component of a continuous speech recognition system, which has a number of uses. Entropy can be used to measure the amount of information added by grammatical and semantic constraints and hence to evaluate different language models. It can also be used as a system internal measure of the confidence that the system has in its answers, which is strongly correlated to the degree of accuracy achieved by the recogniser. These uses are illustrated by data from the ISTD speech recogniser.