Punctuation Generation Integrating Audio Pauses Through Bi-Directional LSTM for Video Lectures

Punctuation prediction plays quite a pivotal role in linguistic communication and communication may face many challenges in its absence. Though there are various traditional models for punctuation prediction, we aim to use the qualities of Bi-Directional LSTM to refine the accuracy and readability of automated speech recognition-generated transcripts by developing a model capable of predicting punctuation marks. This model predicts the punctuation symbol by working in back and forth direction of the text. This work utilizes a two-stage approach: the first stage involves tagging training and validation datasets with explicit punctuation annotations, while the second stage introduces the manual insertion of pauses to refine the model's training data. By training a bidirectional LSTM model, this paper seeks to improve the model's ability to accurately predict punctuation marks like commas, and periods. It was also found that adding pauses in the sentences, enhances the performance of the model in terms of F1 score and with lower error rate and better perplexity for video lectures.