Recognition of Spoken Spelled Names for Directory Assistance Using Speaker-Independent Templates

In the past few years, a great deal has been learned about the processes of automatic speech recognition. As a result, it is now practical to implement isolated word recognizers which achieve high accuracy for a variety of talkers for which the system is trained. 1 For the case of word recognizers that are speaker-independent, we are only now learning how to reliably implement such systems based on statistical characterizations of the variability in speaking the words of the 571 vocabulary.2"5 Because of the success of these recognition efforts, work has been proceeding on task-oriented applications of isolated word recognition. One such example of a task-oriented recognizer is the directory assistance system proposed by Rosenberg and Schmidt. 6 For this system, an isolated word recognizer is linked to a post-processing directory search algorithm to find a name in the directory that best matches the recognizer output (which is, of course, an estimate of the spoken name). A block diagram of this system is given in Fig. 1. The input to the system is a string of isolated words consisting of the letters of the last name (up to six letters), followed by the command word STOP, followed by one or more initials, and a final STOP (to delimit the end of the string). The isolated word recognizer provides a set of candidates for each spoken word (as shown in Fig. 1 for the name LEE K) ordered by recognition distance. The matrix of recognition candidates is used by a post-processor directory search procedure which queries a directory of names to find a best match to the candidate string.