Classroom speech intelligibility prediction using Elman neural network

Abstract

A study was conducted to develop a simple system for classrooms speech intelligibility prediction. In this study, several classrooms properties such as size, signal-to-noise ratio (SNR) and Speech Transmission Index (STI) were collected from different types of classrooms in Universiti Malaysia Perlis (UniMAP). A dataset was obtained from the measurement and was used to develop the system. To develop the system, several process were implemented which includes the statistical analysis, data cleaning and preprocessing, network development, training and classification. In this study, Elman network was selected to develop the system for its robustness in prediction application. The study has also experiments with the network dependency on normalization by comparing different types of normalization method. A simple system for classroom speech intelligibility prediction was developed and it was concluded that network performances are dependent to the normalization method.