Categorising [s], [integral of] and intermediate electropalatographic patterns: neural networks and other approaches

We consider two approaches to the interpretation of EPG data, using materials designed for the investigation of [s]-[integral of] assimilations across word boundaries, as in boss Shep. A contact index is based on electrode contact in different row groups, whereas a neural network is used to determine patterns of contact. Both approaches are speaker sensitive, reflecting the wide variation in [s] and [integral of] articulation between speakers. The advantage of the neural net analysis is that the investigator need not make any assumptions about the relative value of specific electrodes or groups of electrodes. We found that the 'hand-selected' index, in general, modelled linguopalatal movement for all degrees of [s]-[integral of] assimilation better than the neural net. However, the neural net very successfully characterised stable /s/ and /integral of/ and has the potential to give an accurate representation of intermediate assimilatory cases, if trained on a more varied input set than the rather limited one used in this experiment.