The paper introduces a new class of signal processing techniques that are trained on large amounts of speech data to extract a set of features for automatic recognition of speech. Such optimized signal processing techniques appear to be consistent with some important properties of human hearing.