The paper proposes and discusses a machine approach for identification of unexpected (zero or low probability) words. The approach is based on use of two parallel recognition channels, one channel employing sensory information from the speech signal together with a prior context information provided by the pronunciation dictionary and grammatical constraints, to estimate `in-context' posterior probabilities of phonemes, the other channel being independent of the context information and entirely driven by the sensory data to deliver estimates of `out-of-context' posterior probabilities of phonemes. A significant mismatch between the information from these two channels indicates unexpected word. The viability of this concept is demonstrated on identification of out-of-vocabulary digits in continuous digit streams. The comparison of these two channels provides a confidence measure on the output of the recognizer. Unlike conventional confidence measures, this measure is not relying on phone and word segmentation (boundary detection), thus it is not affected by possibly imperfect segment boundary detection. In addition, being a relative measure, it is more discriminative than the conventional posterior based measures.