Rejection in speech recognition systems with limited training

In this paper, we propose a new rejection criterion applicable specifically to limited-training speech recognition systems such as Speaker-Dependent (SD) recognition systems. The new criterion uses confidence measures as well as speakerspecific out-of-vocabulary (OOV) models. The OOV models are created from the same training data that is available to create the in-vocabulary (IV) word models. We describe the method for creating these speaker-specific out-of-vocabulary models from limited training data. We also define a fairly robust confidence measure to reject the OOV words. The results presented in this paper demonstrate the effectiveness of the new criterion in a SD recognition task under various conditions.