Discovering acoustic features with a connectionist learning algorithm
Jeffrey L. Elman, David Zipser · The Journal of the Acoustical Society of America · 1986
A method for discovering acoustic features in speech using a parallel distributed processing (or “connectionist”) learning algorithm is described. The approach involves representing speech patterns as distributed patterns of activation over a network of interconnected processing elements. The “generalized delta rule” (Rumelhart, Hinton, and Williams, 1986) is used to teach the model to autoassociate various speech patterns with themselves. Teaching is accomplished by adjusting connection strengths between processing elements in order to minimize the error in autoassociations. This method was used to discover a set of acoustical features which efficiently and accurately encode speech sounds. The results of several simulations are reported, and the implications of the approach for models of human speech perception, as well as possible applications for machine-based speech recognition systems are discussed. [Work supported by the Office of Naval Research.]