Recognition of spoken spelled names applied to directory assistance

Abstract

The automatic word recognition system originated by Itakura has been modified to accept strings of spelled letters spoken in isolation. The output of the recognizer is a set of best candidate letters for each letter spoken in the string. Candidate strings forming spelled last names and initials are compared to name strings extracted from a telephone directory stored in a disk file. A systematic search is carried out to find a matching entry in the directory. The outcome of a search is one or more matches or no match. An evaluation of the system was carried out using the 17 000 entry Bell Labs telephone directory. A list of 50 randomly selected names was extracted from the directory. Ten talkers participated in the evaluation spelling out each name in the list over a dialed-up telephone line input to the system. In the initial input operation, the median error rate in acoustic recognition of the individual spoken letters may be as high as 20%, but the constraints imposed by the system in making the directory search reduce the error to less than 6% in the final identification of the name.