This paper explores the capability of deep learning to generate lyrics for a designated musical genre. Previous research in the field of computational linguistics has focused on lyric generation for specific genres, limited to Recurrent Neural Networks (RNN) or Gated Recurrent Units (GRU). Instead, we employ a Long Short Term Memory (LSTM) network to produce lyrics for a specific genre given an input sample lyric. In addition, we evaluate our generated lyrics via several linguistic metrics and compare these metrics to those of other genres and to the training set to assess linguistic similarities, differences, and the performance of our network in generating semantically similar lyrics to corresponding genres. We find our LSTM model to generate both rap and pop lyrics well, capturing average line length, and in-song and across-genre word variation very closely to the text it was trained upon.
Gill, Harrison; Lee, Daniel; and Marwell, Nick
"Deep Learning in Musical Lyric Generation: An LSTM-Based Approach,"
The Yale Undergraduate Research Journal: Vol. 1:
1, Article 1.
Available at: https://elischolar.library.yale.edu/yurj/vol1/iss1/1