Automated Generation of Sinhala Lyrics using Recurrent Neural Networks

Abstract

Abstract: This paper discusses the approaches involved in implementing automated song lyrics system in the Sinhala language. Which includes an overview of the complexity of writing song lyrics and develop an automated application for Sinhala song lyrics generation. Before the implementation was carried out, a set of Sinhala song lyrics has been collected to create a corpus, and it has been used to develop an RNN model with LSTM layers using different temperatures and epochs. Then the created models were used to carry out a comparison process to evaluate the effect of the corpus size and the number of epochs per model training to get a better understanding of the RNN training behaviors. Finally, the system was served to a web host to give the user a friendly UI, where the user can enter desired keywords and generate new Sinhala song lyrics. The initial results were obtained through different models and we could see that with the increment of the number of epochs and the number of song lyrics that are trained in each model, the generated output had a clear growth in terms of accuracy and meaning of the song.