Neural network based integration of multiple confidence measures for OOV detection

In this paper we present a novel method to reject OOV words for speaker dependent dynamic command set recognition. The OOV rejection problem is regarded as the designing of recognizer with two classes: In-Vocabulary command and OOV command. Multiple soundly confidence measures derived from likelihood score of acoustic match and prosody match are defined and compete with each other at the same level automatically within neural network framework, thus elude choosing balanced sensitive threshold like traditional strategy. The network weights are trained according to Minimum Misclassification Error criterion. The confidence measures take whole command set into account, and objectively describe the difference between the top one and alternative hypotheses. Experimental results show that neural network based combination is rational, reliable and stable with average total error rates 9.3%, outperforming any single confidence measure threshold approach. Also the across verification results show that trained network is independent of speaker, gender and command set. Although there is performance degradation when exported to another conditions, it is acceptable in many applications.