Context-modulated discrimination of similar vowels using second-order connectionist networks

Abstract

Discrimination of medial adjacent vowels in the context of voiced and unvoiced stop consonants using connectionist networks is investigated. Separate discrimination networks were generated for one speaker from samples of the vowel centers of [e,æ] for the six contexts [b,d,g,p,t,k]. A single context-independent network was similarly generated. The context-specific error rate was 1%, whereas the context-independent error was 9%. A method for merging isomorphic networks into a single network is described. The method uses singular value decomposition to find a minimal basis for the set of context-specific weight vectors. Context-dependent linear combinations of the basis vectors may then be computed using second-order network units. Compact networks can thus be obtained in which the vowel discrimination surfaces are modulated by the phonetic context. In a preliminary experiment, as the number of basis vectors was reduced from 6 to 3, the error rate increased from 1% to 3%. Experiments with nonlinear optimization of context-modulated, second-order networks for this task are underway and will be reported.