A Segmental k-Means Training Procedure for Connected Word Recognition

Abstract

Algorithms for recognizing strings of connected words from whole-word patterns have become highly efficient and accurate, although computation rates remain high. Even the most ambitious connected-word recognition task is practical with today's integrated circuit technology, but extracting reliable, robust whole-word reference patterns still is difficult. In the past, connected-word recognizers relied on isolated-word reference patterns or patterns derived from a limited context (e.g., the middle digit from strings of three digits). These whole-word patterns were adequate for slow rates of articulated speech, but not for strings of words spoken at high rates (e.g., about 200 to 300 words per minute). To alleviate this difficulty, a segmental k-means training procedure was used to extract whole-word patterns from naturally spoken word strings. The segmented words are then used to create a set of word reference patterns for recognition. Recognition string accuracies were 98 to 99 percent for digits in variable length strings and 90 to 98 percent for sentences from an airline reservation task. These performance scores represent significant improvements over previous connected-word recognizers.