This paper presents an architecture for speech understanding that was implemented for an information query application in Italian. The architecture employs speech recognition and stochastic grammar algorithms developed at IRST. Interpretation of an utterance is computed in three steps: first, a phonetic-unit sequence is produced by a speech recognizer with the help of a bigram language model; second, a sequence of semantic units is computed by a Viterbi decoder with the help of a semantic language model; and, finally, a semantic parser computes the most probable semantic interpretation of the query. Experimental results are reported for a test set of 154 queries which were never seen during the architecture design and training.
Bibliographic reference. Federico, Marcello / Vernesoni, Fabrizio (1995): "A speech understanding architecture for an information query system", In EUROSPEECH-1995, 563-566.