Song writing is a mystery to me - being able to find the right words that match the right melodies while also telling a story is a respectable skill. Using Deep Learning models and lyrics from established songwriters, I wanted to attempt to create a text generator for songs.
This is not/will not be used in any commercial capacity, I simply want to find out if I can train a model to make somewhat coherent text.
I repurposed the script from my lyric analysis project, using it to fetch the lyrics of 53 artists. I decided to download 150 songs per artist to begin the training set.
A function fetch_lyrics(artist, retries, max_songs) gathers and saves the lyrics for each artist in the list, using the lyricsgenius library. I delayed this function using dask.delayed, which sped up its run time from over an hour to 21 minutes. The full code is available in lyric_retrieval.py.
The retrieved lyrics were already relatively clean thanks to the lyricsgenius library. Each song came with a header (Song 1, Song 2, etc.), as well as a truncated version of the song's description, taken from the Genius webpage. These are not needed in the model, so I created the function clean_text(txt_file, pairs) which accepts text and a dict of pattern/replacement pairs. The function applies each replacement to each pattern found within the text. The result is one text file with 150 songs per artist, each song separated by a "--next song--" marker.
- figure out how to train model using both structured lyrics + vocab from other places so it is not just limited to the songs.