Analysis of Punctuation Prediction Models for Automated Transcript Generation in MOOC Videos

Abstract

In today’s e-learning based educational scenarios, lot of efforts in terms of time and manpower are required by the MOOC instructors for the generation of transcripts. This research study is focused on the efficient and correct punctuation prediction in the process of automated generation of these transcripts. Various deep learning based and other commonly used punctuation prediction techniques and models existing in the literature have been identified and analyzed for the educational domain videos. The hybrid model of Convolution Neural Networks and Bidirectional Long Short Term Memory ensembled with the acoustic model outperformed other models. It yielded an accuracy of 93.56 percent, recall of 56.15 percent and precision of 63.69 percent. This study also proposed a generalized architecture for efficient punctuation prediction.