Automatic recognition of liquids and nasals in continuous speech

The paper describes the implementation of a Speech Understanding System(SUS) component which emits hypotheses about liquid and nasal consonants. After performing a precategorical classification, the speech message is segmented, under the control of a grammar, into syllabic nuclei and hypotheses about vowels and consonants are emitted with context-dependent procedures. Such procedures are based on a syntax whose rules are inferred by experiments. Furthermore, a semantics is associated to each rule for evaluating each hypothesis by a fuzzy composite relation.