One feature that classification algorithms typically lack is the ability to know what they do not know. With this knowledge an algorithm would be able to operate in any domain and only produce results when it is confident that data is within nominal conditions. Otherwise, it could generate warning messages or request more appropriate training material. We present an unsupervised approach capable of working in concert with an existing classifier to detect off-nominal conditions by estimating the divergence between the distribution of input features and a nominal world model. Using a measure of parametric divergence for a mixture of Gaussians and two different estimates for the Kullback-Leibler divergence, we significantly outperform the baseline average log probability thresholding to distinguish nominal conversational audio from a variety of structured noises and incorrectly decoded audio using features from a speech activity detector.
