Because of the uncertainty involved in terminal recognition, the objective of a syntactic analyzer under uncertainty is no longer to find one structure fitting the speech, but the best one. Two problems must be solved: (1) define and apply an appropriate quality measure of the resulting structures, in order to find the best one, (2) construct and maintain in parallel a set of plausible partial structures, in order to be able to compare them. We present a syntax analyzer which uses the notion of confidence islands realizing: parallel construction of the syntactic structures of a speech signal using beam-search strategy based on a specific quality measure we introduced, mixed bottom-up and top-down extension of parsing trees, and unification of islands relating to different parsing trees. Applied to continuous speech understanding, for a 200 sentence single speaker test, the system obtained a correct rate of 90% at sentence level.