Sentence Based Discourse Classification for Hindi Story Text-to-Speech (TTS) System

In this work, we have proposed an automatic discourse prediction model. It predicts the discourse information for a sentence. In this study, three discourse modes considered are descriptive, narrative and dialogue. The proposed model is developed using story corpus. The story corpus comprises of audio and its corresponding text transcription of short children stories. The development of this model entails two phases: feature extraction and classification of the discourse. The feature extraction is carried out using ‘Word2Vec’ model. The classification of discourse at sentence-level is explored by using Support Vector Machines (SVM), Convolutional Neural Network (CNN) and a combination of CNN-SVM. The main focus of this study is on the usage of CNN for developing the model because it has not been explored much for the problems related to text classification. Experiments are carried out to find the best model parameters (such as the number of the filter, filter-height, cross-validation number, dropout rate, and batch-size) for the CNN. The proposed model achieves its best accuracy 72.6% when support vector machine (SVM) is used for classification and features are extracted from CNN (which is trained using the word2vec feature). This model can leverage the utilization of the discourse as a suprasegmental feature from the perspective of speech.