Evaluation of a Directory Service Search Algorithm

Abstract

To determine the effectiveness of Touch-Tone or ASR (Automatic Speech Recognition) Directory Services, three studies were conducted. Specifically, Roe and Salahi's (1984) search algorithm, based on a Hidden Markov Probability Model, was evaluated. The initial study assessed the algorithm's ability to locate a targetted directory listing from correctly or incorrectly spelled TT inputs. The second study determined the generalizability of the original study, and the third study evaluated the accuracy of the algorithm for voiced-in listings that were interpreted by a speech recognizer. The names were sampled from a 1,450-name AT&T Information Systems data base. For TT inputs, results from the first two studies were similar; overall, the algorithm located 93.5% of the targetted names despite the fact that two thirds of the names were purposely spelled incorrectly. 100% of the correctly spelled names were located. For voice inputs, results showed that the algorithm located 89.9% of the targetted names. These results attest to the fact that this algorithm is highly effective at locating targetted directory listings regardless of spelling errors. The high degree of accuracy is especially impressive in light of the fact that each TT key is associated with 3 or 4 letters of the alphabet and that each voice input may be interpreted as any of the 26 letters of the alphabet.