A hybrid approach to reliable word recognition under adverse environmental conditions is described in this paper. The overall strategy is to find at first a distortion-robust signal representation and to remove the noise components in this domain afterwards. Several preprocessing steps help to enhance the robustness of the basic feature set generated from standard lpc-cepstrum analysis: calculation of temporal derivatives, principal component analysis and reordering of the coefficient set with respect to a predefined figure of merit. Subsequent neural network-based noise reduction widely removes the remaining noise component. The contribution of each step is evaluated separately by an isolated word recognizer, and some aspects of cost reduction for realtime implementation are discussed.
Keywords: dimensionality reduction, feature vectors, Lombard effect, neural networks, noise reduction, noise robustness, word recognition