Recognition of spoken spelled names for directory assistance using speaker-independent templates

In a recent paper, Rosenberg and Schmidt demonstrated the applicability of a speaker-trained, isolated word speech recognizer to the problem of automatic directory assistance. Input to the system was in the form of a string of letters which spelled the last name and initials of an individual for whom a directory listing was required. Rosenberg and Schmidt found that, even though the recognition rate for individual letters was rather low (approximately 80 percent), the rate at which the correct directory listing was found was higher (approximately 95 percent). In this paper, we extend these results to include the case of speaker-independent recognition of letters. We show that overall performance in the speaker-independent mode is comparable to performance in a speaker-dependent mode and examine various factors important for operation in a speaker-independent mode, such as characteristics of the reference templates, choice of decision rule, and threshold parameters. For the most part, the overall system is remarkably robust to the parameters of the recognizer. For the best choice of these parameters, a 95-percent correct string rate is obtained, comparable to the performance in a speaker-dependent mode.