Learning to Segment Speech with Self-Organising Maps

Abstract

In recent years, a number of models of speech segmentation have been developed, including models based on artificial neural networks (ANNs). The latter involved training a recurrent network to predict the next phoneme or utterance boundary, and deriving a means of predicting word boundaries from its behaviour. Here, a different connectionist approach to the task is investigated employing self-organising maps (SOMs) (Kohonen 1990). SOMs differ from other ANNs in that they are unsupervised learners. The aim is to investigate whether the SOM can become sensitive to where word boundaries occur, when trained on phonetically transcribed speech.