A duration-based confidence measure for automatic segmentation of noise corrupted speech

In this study, a duration{based measure is formulated for assigning con dence scores to phonetic time{alignments produced by an automatic speech segmentation system. For speech corrupted by additive noise or telephone channel environments, the proposed con dence measure is shown to provide a reliable means by which gross segmentation errors can be automatically detected and marked for human hand correction. The measure is evaluated by computing Receiver Operating Characteristic (ROC) curves to illustrate the expected trade{o in probability of detecting gross segmentation errors versus false alarm rates.