The purpose of this work is two-folded: to improve the robustness of a baseline isolated-word recogniser with a medium-sized vocabulary in terms of word rejection capabilities and to verify these improvements on a multi-application database including different native and non-native English accents. Although the results obtained so far could be biased by the unavailability of realistic corpora, they seem to indicate the usefulness of multiple sink models in this context, relative to the use of a single one. The improvements are particularly evident in multi-accent environments, where our results show that merging two different accents in the training material yields scores which are similar to the ones obtained in a single accent environment.