While the phoneme is generally accepted as the basic recognition unit
in large vocabulary isolated word automatic speech recognition, in
continuous speech recognition phonetic target undershoot often leeds to
poor phoneme definition. It is now established that phoneme transitions
can provide a complementary set of speech units which are less prone to
the effects of context and coarticulation, but as yet we do not have a
clear understanding of the phonetic invariants which these transitions
provide. In this talk I will describe and present the results from
some human perception and machine recognition experiments which we have
designed specifically to analyse the phonetically relevent information
for consonantal place of articulation in a multilingual set of VC and CV
transitions.