As part of an effort to do high-accuracy speaker-independent recognition of the English alphabet, we focus our attention on the E-set, a very difficult subset of nine letters: B,C,D,E,G,P,T,V,Z. By adding knowledge-based features and finding the most suitable spectral representation, we are able to reduce the E-set error rate by 20% compared to a baseline system that uses features designed for the whole alphabet. The resulting E-set classifier was successfully added to the full-alphabet system, to be invoked whenever the first answer is any member of the E-set.