Introduction

At the peak of the influence of Generative Phonology, Charles Ferguson and Carol Farwell of the Stanford Child Phonology Project published a troubling observation (Ferguson and Farwell 1975), based on longitudinal data: they showed that the construct of ‘phoneme’ did not (indeed, does not) do justice to the patterns of variation in some young children's production of speech sounds (see also Menyuk et al. 1986). But data which, like these, do not fit any recognisable theory are like a strange tool that comes without instructions, or an unfamiliar spice for which one has no recipe. They stay in a box, untouched. Perhaps they are retrieved from storage when some use for them comes along; but more likely, they are only noticed lurking there after something similar has been rediscovered somewhere else, at a time and place where the odd item can at last be assimilated (cf. the rediscovery of Mendelian heredity).

Once we do have a new beautiful theory that can handle data which were previously intractable, we tend to go on a binge: we try to show that it can do almost everything. That is good – we have to check out its power. But we also tend to shove all the data that do not fit the new theory into that storage box, whether or not they were well handled by a previous approach.