Musician and programmer Sam Agnew needed some fresh quar tunes so badly, that he created a lyric generation bot. The program leverages OpenAI’s Generative Pre-trained Transformer 3 (GPT-3) which has been trained on massive amounts of text. By texting an artist’s name to +1 (315) 65-LYRICS (+1 315 659-7427) in the U.S. or +44 740 119-3427 in the U.K., you’ll quickly receive some lyrics in their style. Singer-songwriter types tend to yield better results while pop stars can get drowned in vocalizations and platitudes.
What to expect — The less info the GPT-3 has on an artist, the shorter the lyric response will likely be. This is most likely to affect indie artists, but R&B goddess Jill Scott tends to fall into this category. Hip-hop, R&B, and rap artists generally get mixed results, from lifting full Beyoncé lyrics to dropping song titles (feature list and all) into Cardi B lyrics. R&B-influenced, genre-straddling Billie Eilish confused the AI significantly:
Despite BTS being one of, if not the, most popular artists in the world, their generated Korean lyrics often make little sense. It’s not clear if this has more to do with the language or that their catchy hits hinge on repetition. Rising, Spanish-singing star Rosalía, however, should probably record this:
Just you wait for GPT-4 — The GPT-3 is imperfect, but increasingly impressive — as demonstrated by the sneaky blogs and creepy photos created using it. The mixed results for different artists in Agnew’s text bot tells a strong story about exactly what kind of text this AI was fed. An attempt for Marvin Gaye simply returned a mini-biography, but Bob Dylan returns nearly flawless lyrics.
So far, many tests of this bot on social media have leaned towards genres like indie folk, rock, and even metal with positive results, but the AI struggles with the most popular genres on the Billboard 100 right now. Even so, here I am, texting “Rihanna” every couple of minutes just to get new lyrics, no matter how terrible.