Some applications of second-order connectionist networks to speech recognition problems

Second-order connectionist networks have been identified as good models for representing acoustic phonetic invariance, since they can modulate separating hypersurfaces or transform data representations as a function of context. These capabilities are illustrated for two problems in vowel recognition: speaker normalization and phonetic context dependency. The idea of context dependency can also be extended to the notion of state in recurrent networks. Second-order recurrent networks that recognize simple finite state languages over {0,1}^* have been induced from positive and negative examples. Some implications of these results for recognizing phoneme sequences are discussed