ENHANCING TEACHER IDENTIFICATION SYSTEMS WITH SUBJECT RECOGNITION USING AUDIO-BASED DEEP LEARNING
Abstract
Teacher identification systems have gained prominence in the educational technology research because of the growing need to use automated academic monitoring and institutional accountability. This research paper designed and tested an audio-based deep learning model that can identify teachers and at the same time detect the subjects being taught by the teacher in the classroom. The researchers used a mixed-method experimental design and gathered about 600 hours of labeled audio of 120 teachers in 10 academic subjects in five secondary schools. It was developed as a multi-task deep learning architecture, which consists of a Convolutional Neural Network (CNN) to extract spectrogram-based features and a Bidirectional Long Short-Term Memory (BiLSTM) layer to model the time. Two output lines were set up to do parallel classification of speaker identity and subject domain. Spectrogram representations, pitch contour features, and Mel-frequency cepstral coefficient (MFCCs) were obtained after noise reduction based on spectral subtraction. The CNN-BiLSTM model proposed had the highest overall accuracy of 96.8, which is much higher than the traditional baseline classifiers, such as Support Vector Machines (SVM) with 78.3% and Gaussian Mixture Models (GMM) with 74.5%. The results of the subjects showed that the classification performance was high in all ten disciplines. The results showed that the audio-based deep learning systems are a valid, scalable, and non-invasive method of improving teacher recognition in contemporary learning institutions.
Keywords : Teacher identification systems, educational technology, automated academic monitoring, institutional accountability, deep learning model, Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM).
