Nowadays automatic speech recognition systems rely on stochastical models, like HMMs (Hidden Markov Models) or ANNs (Artificial Neural Networks). The featurally underspecified lexicon (FUL) system uses a radically underspecified phonological representation as basis for the recognition process. Monovalent features are extracted from the speech stream and are mapped with a ternary logic (match, no-mismatch, mismatch) to the entries in the 50 000 lemma-based word lexicon. This mapping is performed without building a segmental representation or any other structures like e.g. syllables. The mapping process operates from left-to-right and allows for certain assimilatory modifications of segments of the words in the lexicon as they occur in running speech. Matching features increase the score of the lexical entry, no-mismatching features do not exclude word candidates, and only mismatching features deactivate words from the search. Candidates remaining from this selection form sequences of word candidates that undergo a parallel active prosodic and syntactic parsing of phrase hypothesis. The intention of the system is in the first place not to build a commercial speech recognition system but to use it as a tool to test hypotheses of phonological representation and subsequent syntactic parsing.