GENERATIVE AI POWERED LEARNING COMPANION FOR PERSONALISED EDUCATION AND BROADER ACCESSIBILITY

This research presents the development and evaluation of a hybrid Convolutional Neural Network (CNN) and the Bidirectional long -term short -term memory (BILSTM) model for speech recognition, especially tailored for educational applications.Using the Mozilla Common Voice Dataset, the model suffered an impressive testing accuracy of 91.87% and less testing loss of 0.2966.The study highlighted the importance of effective preprocessing, including noise reduction, audio trimming, and MEL-Frequency Cepstral Coefficients (MFCC) feature extraction, which were necessary to improve model performance.The CNN-BiLSTM architecture enabled the model to capture both local and longrange temporary dependence, making it strong for diverse accents, speech speeds and background noise.This task reflects the viability of implementing advanced speech recognition systems in the generative AI-in-charge learners, contributing to the manufacture of inclusive and accessible educational devices.Future research can detect fine-tuning for specific domains to carry forward multilingual dataset, attention mechanisms, and performance.