Convolutional Neural Network Based Broadcast News Summarization using Acoustic-Prosodic Features

Speech summarization is aimed at reducing the length of spoken documents while preserving the core information. Extractive summarization creates a summary by selecting the important sentences. This paper proposes a novel one-dimensional convolutional neural network (1D-CNN) architecture, for summarizing broadcast news. The extractive speech summarization task is framed as a binary classification problem, where the model processes spoken sentences segmented from the spoken doc-ument. The proposed 1D-CNN model is trained on handcrafted acoustic-prosodic feature representation of the spoken sentences, thus making this approach text-independent. This is particularly relevant for low-resource languages as this allows effective sum-marization with limited training data, and not being dependent on speech converted text transcripts. The experiments in this paper were conducted on broad-cast news data in a low-resource language, and the summaries are evaluated using the ROUGE metric.