HARPY
Lowerre's system recognised 1011 words of continuous speech by compiling every level of knowledge into one large transition network and searching it with a beam of best paths.
Why it matters
The five-year programme's goal was met, but at a cost that showed the approach's limit: the network had to be built in advance for each task.
HARPY combined ideas from HEARSAY-I and DRAGON: a deterministic state network without prior probabilities, traversed by beam search. Sentence accuracy ran from 70 to 100 percent depending on the task, at one and a half to ten times real time. The network was compiled separately for each new domain and held millions of states, so the approach did not transfer.