Experimental Investigation into Alignment-based Acoustic Confidence Measures in Keyword Verification for Mandarin Speech

This paper introduces the methods to verify keyword hypothesis using alignment information, which can be easily implemented and integrated into the keyword spotting system in a straightforward way. Acquiring alignment information goes without any extra model training, and has no evident total processing time impact on original keyword spotting system. Different alignment-based confidence measures are evaluate using ROC curves. In addition, we introduce a novel evaluation metric, Confidence Error Cost Function (CECF), which improves upon Confidence Error Rate (CER). It is more general than CER since it takes different cost weights of FA and FR into consideration. We present the result of experiments aiming at assessing the quality and the limitations of different confidence measures and find out that the combination of minimum alignment cost and state duration normalized posterior probability gives the best performance.