Lyrics Generation Using LSTM and RNN

Abstract

For years, both the music and AI industries have been interested in automatic lyric creation. Early rule-based techniques have mostly been supplanted by deep-learning-based systems as computing power and data-driven models have evolved. This paper explores the capability of deep learning models to generate lyrics for a designated musical genre and artist. Previous lyric generation research in the field of computational linguistics has been restricted to Recurrent Neural Networks (RNN) or Gated Recurrent Units (GRU). On the contrary, this paper focuses on the use of LSTM networks to generate lyrics for a certain genre/artist from a sample lyric (seed) as input. The study also looks at how modern deep learning techniques may be used to improve the songwriting process by learning from artists and their significant works. The LSTM model proposed by this paper was able to generate both raps as well as pop lyrics, capturing average line length, in-song, and across-genre word variation very closely to the text it was trained upon. By adjusting the model's parameters, the neural network was pre-trained on the lyrics of different composers and musicians. A multilayer LSTM-based training model with bidirectional neurons and BERT integration is used in this study. To generate a comprehensive set of lyrics, the lyrics supplied as input are divided down into a word and rhyme index. This model is generative; therefore, each trial yielded a unique set of lyrics. The model produces a loss of less than 0.06 when the parameters are set correctly. The differences in the results of permutations of different dropout positions were analyzed. Some of the model's lyrics were determined to be of suitable quality. To summarize, the findings suggest that deep learning techniques can be utilized to support the creative process of songwriting.