Analyses of the hidden units of the multi-layer perceptron and its application in acoustic-to-articulatory mapping

Abstract

An artificial neural network (ANN) is applied to perform the task of acoustic-to-articulatory inversion. The objective is to model the highly nonlinear mapping from linear predictive coding (LPC) code to corresponding articulatory parameters with a multilayer perceptron ANN structure. Such information will facilitate the study of the relationships between the acoustic signal and the physical vocal tract which produces it. Several novel approaches for devising the ANN structure have been evaluated. Specifically, the performance of two learning algorithms, a backpropagation (BP) algorithm, and a random optimization (RM) algorithms, are compared. To reduce excessive, redundant hidden units in the multilayer perceptron model, a singular value decomposition is applied to either the weight matrix or the output covariance matrix of the hidden units to check their corresponding ranks. In both cases, their ranks are closely related to the number of essential decision regions in the input data. >