Autoassociative neural network for speech processing
Procheta Poddar, P. V. Rao · International Conference on Artificial Neural Networks · 1991
Connectionist architectures are being studied from various perspectives and applied in diverse domains with promising performance. In the paper, the authors study MultiLayer Perceptron (MLP), one of the most widely used architectures, for generating alternative representations of speech signal. Speech is a highly redundant signal and hence efficient representation of speech signal that exploits this redundancy is an important issue in synthesis, transmission and recognition of the human voice. It has been observed that MLP forms suitable internal representations in terms of the activation of its units to establish an association as specified by a given input-output relation. The authors explore the nature of this internal representation formed by an MLP while establishing an autoassociation of spectral patterns of speech signal. >