This research is aimed at the development of a speaker- independent phonetic classification system for continuous speech which retains the statistical and automatic-training capabilities of implicit representational strategies, while accommodating explicit and intuitive descriptions of recognition vocabularies. The classifier is being implemented within a recognition framework with three main components: a finite-state pronunciation network whose branches correspond to linguistically well-defined units; a set of generalized acoustic pattern matchers; and a segment-based dynamic programming search. A primary feature of this framework is its ability to incorporate acoustic-phonetic features computed over variable-duration, segment-sized units of the speech signal, in addition to the general fixed-rate acoustic measures typically employed in traditional pattern-matching systems.