Classroom speech intelligibility prediction system for front-rear speech amplified classroom based on audio features

Abstract

Classroom speech intelligibility is a measure of how well a speech or word is understood in the classroom. It is a measure of the speech quality in the classroom. Numbers of methods have been proposed by various researchers to improve the speech intelligibility. However, the proposed methods are effective only in the design stage of the classroom, as implementation in the ‘old’ classroom is costly and time consuming. Thus, speech amplification is implemented to tackle such problems. There are methods suggested by audio expert on how to properly setup the system in the classroom, in order to maximize the speech intelligibility. However, the methods are rather complicated and time consuming. So, as an alternative, this research has proposed an audio-feature based speech intelligibility prediction system. The goal of this research is to develop an intelligent speech intelligibility prediction system by combining audio-features (spectral rolloff (SR), spectral centroid (SC), power (PO), zero-crossings rate (ZCR), and short time energy (STE)) and classifiers (feed forward neural network (FFNN), Elman network (ENN)). To achieve the goal, this research has collected data samples which comprises of speech recordings in the speech amplified classrooms, as well as the physical properties. The measurement was done in eight different classrooms in UniMAP, and the measurement protocol was derived from the previous researches and acoustic standards. The data collected were then analyzed using statistical approach, such as descriptive analysis and ANOVA. The data were then pre-processed to assist the later feature extraction process. The preprocessed signals were then undergone feature extraction process to extract the audio features. In this research, five types of audio features have been selected, and each feature is then combined with the classroom’s physical feature data as inputs of the experimented classifiers. As a result, it was found that audio feature PO yield the best accuracy, regardless the type of classifiers when compared to the other features. At the end, the interface system for the audio feature-based classroom speech intelligibility prediction system is developed. Moreover, a database of classroom speech intelligibility measurement using single microphone was compiled.