No abstract

Automatic Speech Recognition (ASR) technology has the potential to improve the learning experience of students in the classroom. This article addresses some of the key theoretical areas identified in the pursuit of implementing a speech recognition system, capable of lesson summary generation in the educational setting. The article discusses: some of the applications of ASR technology in education; prominent feature extraction and speech enhancement techniques typically applied to digital speech; and established neural network-based machine learning models capable of keyword spotting or continuous speech recognition. Following the theoretical investigation, a model is proposed for the implementation of an automatic speech recognition system in a noisy educational environment to facilitate automated, speech-driven lesson summary generation. A prototype system was developed and improved based on this model, ultimately proving itself capable of generating a lesson summary intended to bolster students' secondary contact with lesson content. This topic-oriented lesson summary provides students with a lesson transcript, but also helps them to monitor educator-defined keyword terms, their prevalence and order as communicated in the lesson, and their associations with educatordefined sections of course content. The prototype was developed using the Python programming language with a modular approach so that its implemented Continuous Speech Recognition system and noise management technique could be chosen at run-time. The prototype contrasts the performance of CMUSphinx and Google Speech Recognition for ASR, both accessed via a cloud-based programming library, and compared the change in accuracy when applying noise injection, noise cancellation or noise reduction to the educator's speech. Proof of concept was established using the Google Speech Recognition System, which prevailed over CMUSphinx and enabled the prototype to achieve 100,00% accuracy in keyword identification and association on noise-free speech, contrasted with a 96,93% accuracy in keyword identification and association on noise-polluted speech using a noise-cancellation technique.