Syllable recognition using hierarchical self-organizing maps.

Abstract

A self-organizing neural network model that is able to learn and recognize temporal pattern sequences in speech such as syllables is described. This model has a hierarchical structure and consists of three parts: a feature-detecting part (feature map), a phoneme-detecting part (phoneme map), and a syllable-detecting part (syllable map). The feature-detecting part is composed of three feature maps which represent static, dynamic, and global feature of a short-time speech segment, respectively. The phoneme map is to identify a phonemelike unit to which each short-time segment belongs, using the response patterns of the three feature maps. The syllable map tries to recognize a time sequence of response units in the phoneme map as a syllable. The time sequence may be represented in a single map by superimposing the response pattern at time t on a decayed pattern of the pattern at t−1. It is shown that the proposed network can be self-organized using the Kohonen learning rule and it works reasonably through phoneme and syllable recognition experiments.