Speech understanding for document retrieval [CMU, 1980]. Uses a precompiled network knowledge structure. The net is searched using beam search.
The HARPY system was developed at CMU after extensive evaluation of HEARSAY-I and DRAGON. HARPY uses a single precompiled network knowledge structure. The network contains knowledge at all levels: acoustic, phonemic, lexical, syntactic, and semantic. The network contains all possible legal sentences in an efficient representation (though it also represents nonsense sentences as a result).
HARPY does a beam search so it never needs to backtrack, and only extends paths in the beam as long as they retain sufficiently high scores. HARPY was the most accurate and efficient performer of the all the ARPA SUR systems.
The system is very sensitive to missing or poorly-labeled segments since the network is composed of a sequential phonetic representation. Design cannot easily accommodate the pragmatics of the utterance, which other systems use to constrain search.