We describe a language model suitable for integration with an HMM-based, speaker-independent, 1000-word recognizer. The system accepts spoken queries and commands to an electronic mailbox. The model represents linguistic information through one single finite state network, whose size is kept small thanks to a procedure for ignoring language constructs not semantically relevant for the task. Experimental evaluations show that this approach competes with other approaches based on the use of phrase-structure grammars for parsing lattices of word hypotheses.
Bibliographic reference. Giachin, Egidio P. (1991): "A dynamic programming based framework for stochastic spoken language understanding", In EUROSPEECH-1991, 391-394.