Neural networks for neutrality classification of Filipino call center agents' English pronunciation

Date of Award

2018

Document Type

Thesis

Degree Name

Master of Science in Computer Science

Department

Information Systems & Computer Science

First Advisor

Fernandez, Proceso L., Jr., Ph.D.

Abstract

This study explored methods of designing and training neural networks to automatically classify the neutrality of Filipino call center agents English pronunciation based on their employers standards on speaking proficiency. Using Mel Frequency Cepstral Coefficients (MFCCs) as features for semi-supervised training, the study found that a standard Artificial Neural Network (ANN) and a Recurrent Neural Network with Long Short-Term Memory (RNN-LSTM) can be designed and trained as utterance neutrality classifiers to automatically classify pronunciations as Neutral or Not Neutral-- yielding high accuracy and F1 scores of 98.69% with 0.99 for the RNN-LSTM and 96.99% with 0.98 for the standard ANN. Hence, these classifiers can capture an unbiased, objective standard of pronunciation specific to the call center involved using the studys methodology. The study also showed that neural networks can produce excellent results on speech classification tasks despite having a small dataset of 380 utterances.

Comments

The C7.B368 2018

Link to Full Text

Share

COinS