Artificial Neural Network (ANN) in a small dataset to determine neutrality in the pronunciation of English as a foreign language in Filipino call center agents

Abstract

Artificial Neural Networks (ANNs) have continued to be efficient models in solving classification problems. In this paper, we explore the use of an ANN with a small dataset to accurately classify whether Filipino call center agents’ pronunciations are neutral or not based on their employer’s standards. Isolated utterances of the ten most commonly used words in the call center were recorded from eleven agents creating a dataset of 110 utterances. Two learning specialists were consulted to establish ground truths and Cohen’s Kappa was computed as 0.82, validating the reliability of the dataset. The first thirteen Mel-Frequency Cepstral Coefficients (MFCCs) were then extracted from each word and an ANN was trained with Ten-fold Stratified Cross Validation. Experimental results on the model recorded a classification accuracy of 89.60% supported by an overall F-Score of 0.92.