A Minimum-Distance Search Technique and Its Application to Automatic Directory Assistance

Automatic methods for obtaining telephone directory information from spoken spelled input can be highly useful. This directory assistance task, however, is one of considerable difficulty in which the problems of automatic speech recognition are exacerbated by the confusability of the vocabulary involved. This is especially true when telephone quality speech is used, since the acoustic differences within certain subsets of the alphabet are extremely small. In particular, such similar-sounding letters as "A," "J," and "K," or "B," "D," "P," and " V " are likely to be confused. In this paper, we consider systems in which the recognition of names takes place in two distinct stages, an acoustic analysis phase, in which the individual letters of the input are recognized, and a search phase, during which a directory is interrogated. We assume that the acoustic 1343 analyzer and its performance are given and focus our attention on devising a search method that allows the correct recognition of spelled names despite the errors introduced during the recognition of the input letters. Recently, Rosenberg and Schmidt1 have described a method which achieves a 92-percent correct recognition rate for names when driven by an acoustic analyzer with an 80-percent correct recognition rate for individual letters; their results were obtained using an 18,000-entry directory. Their method is to look in the directory for each name in a sequence of candidate names constructed from the output of the acoustic analyzer.