On the identification of phonemes using acoustic-phonetic features derived by a self-organising neural network

Abstract

A self-organizing neural network (SONN) is subjected to a training and calibration process using continuous speech spoken by three talkers. The aim of this process is to establish a system which is able to transform speech frame cepstrum vectors into vectors of continuous valued acoustic-phonetic features. The calibration process also involves a stage where each neuron of the SONN is assigned a vector defining the links between speech technology and articulatory phonetic concepts. The validity of the transformation approach is shown by applying a speech test corpus to the SONN transformation. The main results of the established transformation technique are given in a number of histograms by which it is shown that the computed acoustic-phonetic feature values to a large extent are in accordance with the phonological specifications used in the feature transformation. The histograms are further used to demonstrate the ability of the acoustic-phonetic features to identify individual phonemes and to discriminate between vocalic and consonantal phonemes.<>