Acoustic-phonetic attribute determination using multi-layer perceptrons

The multi-layer perceptron (MP) provides a means of performing complex pattern recognition and feature detection tasks. It is capable of implementing non-linear transformations which may be found by means of an iterative training procedure, known at the generalized delta rule. Such a technique is therefore of particular interest in determining elementary phonetic attributes of the speech signal, as first results suggest. This paper describes the application of an MLP network to the feature-labelling of simple speech material. The output of the network is a feature vector for each input time-window. The network can be viewed as either a pre-processing stage for a phonetic recognition system, or simply as a non-linear data reduction of the signal for input to pattern-matching recognisers