Network for the recognition of a large spoken vocabulary

A new solution for accessing a large lexicon in continuous speech is proposed. The words of a lexicon and their relations with syllables, acoustic and prosodic cues are represented by a network similar to semantic networks used in artificial intelligence for knowledge representation. The network for lexical representation is generated by a graph grammar in which rules are applied either when suitable acoustic cues are detected in the signal or when model driven predictions are made. The piece of network generated by the application of a set of rules contains nodes and links. Nodes are processes capable of executing algorithms. Links are channels through which messages between processes can be exchanged. The type of a link specifies the type of message that can be exchanged. The model exhibits a high degree of parallelism and allows to efficiently perform intersections of large sets of descriptions. Some nodes correspond to word classes which are hypothesized when some sets of sufficient conditions are detected in the data. These conditions are hypotheses about robust features called primary phonetic features. 11 references.