Modeling the emergence of bilingual acoustic clusters: a preliminary case study

This paper presents some preliminary results of an original study to model the emergence of bilingual acoustic clusters of both New Zealand English and New Zealand Maori speech. This is performed using true on-line learning in a connectionist architecture. The study represents a joint collaborative analysis, which applies the bilingual data as training examples to a connectionist-based evolving clustering method algorithm. The algorithm returns a structure containing acoustic clusters plotted using visualization techniques that could be used as the foundations for future speech classification systems. The following experiments are based on the notion that approximately 75% of the phonological units in New Zealand English and New Zealand Maori occupy similar acoustic space, they sound the same, and therefore they can be used to classify new unknown speech units or words. (C) 2003 Elsevier Inc. All rights reserved.