Speech recognition issues for directory assistance applications

Abstract

Telephone companies in the United States handle over 6 billion Directory Assistance (DA) calls each year. Automation of even a portion of DA calls could significantly reduce the cost of DA services. This paper explores two factors affecting successful automation of DA: (a) the effect of directory size on speech recognition performance, and (b) the complexity of existing DA call interactions. Speech recognition performance for a set of 200 spoken names was measured for directories ranging from 200 to 1.5 million unique names. Recognition accuracy decreased from 82.5% for a 200-name directory to 16.5% for a 1.5 million name directory. In part because high recognition accuracy is not easily achievable for these very large, low-context directories, it is likely that initial implementations of DA automation will focus on a small percentage of calls, requiring a smaller vocabulary. To maximize potential savings, listings that are most frequently requested constitute the optimal vocabulary. To identify critical issues in automating frequent DA requests, approximately 13,000 DA calls from an office near a major metropolitan area in the United States were studied. In this sample, 245 listings covered 10% of the call volume, and 870 listings covered 20% of the call volume.