Phonetic Classification in Spanish Using a Hierarchy of Specialized ANNs

Abstract

A neural net based methodology for phonetic classification with telephone speech in spanish is described. Because of the high computational requirements and error rates obtained by using a unique Multilayer Perceptron (MLP), a different approach is needed in order to improve the performance of the task.

In the proposed approach, the basic set of spanish phonemes is separated in groups according to articulation mode criteria and a Multilayer Perceptron (MLP) is trained for every phonetic group, along with a front-end MLP whose function is to distinguish between phonetic groups.

Experiments were made with speakers from the telephone speech OGI corpus in order to tune the parameters of the MLPs, as well as to evaluate the performance of the proposed methodology under different representations of the speech signal and modifying some parameters of the ANNs such as learning rate, topology and transfer functions.

Results of the experiments are summarized and some remarks are passed. Both, results and remarks, are based on the analysis of the confusion matrixes obtained when the trained MLPs are used to classify speech used for training as well as speech data that the MLPs haven’t “seen”.

The authors wish to thank the Center for Spoken Language Understanding (CSLU) at the OGI for their kindness in providing the corpus for this worh.