Fuzzy semantic network for a speech understanding system - an experimental study

The relevance of the Semantics for a Speech Understanding System (SUS) was pointed out by many authors and is by now a classic topic. The present work contains an implementation of a semantic analysis for the SUS of the CENS, in Turin. The semantic knowledge is contained in a graph on which a set of fuzzy relations is defined. The semantic algorithm accepts, as input, a table of alternative weighted words, found in the processed spectrogram at phonological and lexical level. The output is a sentence assumed as the "best" interpretation of the whole spectrogram. The algorithm is implemented in FORTRAN and PL1 on a IBM 370/158 and it utilizes a recursive evaluation function for the concurrent hypotheses. We compare experimentally the efficiency of the algorithm, by varying the number of alternative input words and their weight, the number of vertices and arcs of the semantic graph and the evaluation function. The results show a good performance of this pure semantic analysis.