On the use of neural networks for automatic vowel recognition

Abstract

An approach to face the problem of speaker -independent vowel recognition using Artificial Neural Networks (ANN) is presented. The approach lies in algorithms and techniques for training the ANN with Backpropagation rule and then for testing it with the spectral representations of vowels obtained by Linear Predictive Coding (LPC). Speech patterns consist in 12 LPC spectral components obtained from the stationary zones of the vowels and they compose a speech database collected in laboratory conditions with computer noise from a number of speakers. A training set and a testing set of vowel patterns were designed. Network topology is a Multilayer Perceptron (MLP) with a variable number of processing units in the hidden layer. Three types of MLP were tested and the results on vowel recognition rate are presented under this framework.

Speaker-independent Malay vowel recognition of children using multi-layer perceptron

“…Thus, in order to assume that it is stationary, the speech signal needs to be examined in a short segment. For example, the vowels are examined in its stationary region, in a fixed short frame of 25.6 ms [2]. The frame length actually can be variable from 5 ms up to 40 ms.…”

“…The training error and the hidden neuron number of MLP are normally fixed for the training and testing [2,6]. The performance of MLP depends on the appropriate setting of hidden neuron number as well as the training error.…”