Conventional features in automatic recognition of speech describe the instantaneous shape of a short-term spectrum of speech and the pattern classification module is relying on information extracted from large amounts of acoustic and text training data. The article describes an alternative approach where the feature extraction module is trained on data. These data-derived features are consistent with auditory-like frequency resolution and with temporal properties of human hearing. The features describe instantaneous likelihoods of sub-word classes and are derived from temporal trajectories of band-limited spectral densities in the vicinity of the given instant. The paper presents some rationale behind the data-driven approach, briefly describes the technique, points to relevant publications and summarizes results achieved so far.