Unsupervised Learning of Simple Speech Production Based on Soft Competitive Learning

Abstract

In this paper we present a simple connectionist model for the adaptive sensorymotor loop involved in perceiving and producing speech. At the heart of the production part lies an articulatory model which approximates the human vocal tract through polygons and splines. Output of this model is the envelope of the acoustic filter function, realized by this vocal tract, which is comparable to the spectrum of real speech segments. The goal of this research was to find a learning method to train a multi-layer neural network to produce the correct set of twelve articulatory parameters when given the spectrum of recorded real speech (stationary vowels). The method introduced in this paper explicitly makes use of a neural network categorization component. Through so-called soft competitive learning it learns to gradually compress the responses to more and more unitized categorical patterns. After a precategorization phase, during which presented real speech patterns are classified, the model starts to randomly produce output signals. A goodness-of-fit measure, which can be computed easily, is taken as the criterion whether the self-produced signal is close enough to any of the known categories, and as the learning rate to adapt the weights between the categorization layer and the output units.