Anthropomorphic feature extraction algorithm for speech recognition in adverse environments

Speech recognition engines should remain reasonably accurate in adverse environments in order to find their ways from laboratories towards applications. However the human auditory system has been proven to be a versatile tool, which is capable of outperforming the known artificial algorithms in their target environments. Recent advances in psychoacoustics and auditory physiology pointed to the essentially non-linear behaviour of the auditory apparatus. On the basis of the interpretation of the biological information processing it is possible to construct a parametric “human-like” nonlinear algorithm, which exhibit properties similar to those of the live system. Besides the description of the anthropomorphic feature extraction algorithm in this paper we test its performance in accordance with the formulated requirements to the efficient and robust feature extraction and also provide a comparative benchmark of compact ASR system in combination with the proposed algorithm in adverse conditions.