A storyteller's tale: Literature audiobooks genre classification using CNN and RNN architectures

Identifying acoustic properties that characterize reading literary genres can assist in giving a more personal and human tone to the speech of bots and automatic readings. In this paper we consider the following question: given speech segments of audiobooks, how well can we classify them according to their literary genres? In this study we consider three different literary genres: children, horror and suspense, and humorous audio books, taken from two free audio books sites: Librivox and YouTube. We ran four classification experiments: three for each pair of genres, and one for all three genres together. We repeated each experiment twice, with two different network architectures: Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). Note that, throughout the reading, there are sections that are more typical to the book's genre than others. As the samples were taken sequentially throughout the reading of the books and were short in duration, we did not expect high classification rates. Nevertheless, the accuracy of all the experiments were at least 72% for all the pair's classifications; and at least 57% for both architectures for the three classes classifications.