A Training Procedure for a Segment Based Network Approach to Isolated Word Recognition

Even for isolated word recognition, the word-model-based approach is inadequate for certain applications where the performance is limited by possible endpointing errors, by the inefficiency of the word model in representing different pronunciations for the same word or by the weak capability of the model for discriminating minimal word pairs. Without resorting to such remedial procedures as relaxed endpoint dynamic programming, multiple clustered models and a second pass discrimination strategy, the recognition performance can not be improved easily. These problems can be alleviated to a large extent by using a segment-based network. In addition to a potential performance improvement, a segment-based approach can reduce the amount of training data significantly when the segmental units are appropriately chosen. This reduction is especially beneficial for a large vocabulary speech recognition application.