DeepLyrics: Conditional Lyrics Generation with Language Representation Models

Dissecting what it means for a machine to be creative using deep learning to generate song lyrics.

Composing a creative and novel song consists of two main components. First, DeepLyrics learns the rules of language by observing previous bodies of work. We call these rules the model. Besides some initial encoding rules, we allow Deep Learning techniques to implicitly learn language rules on its own such as part-of-speech, grammar, rhyme, and song structure. Furthermore, song lyrics allow for the exploitation of other modes of data — such as those related to the music itself.

The generator is the second component. This is where we achieve the emulation of creativity. While staying within the confines of the model, we add randomness, noise, and blending of existing attributes to generate creative and novel lyrics.

Read more and try DeepLyrics for yourself here!