Say, today, all your music is stored and streamed to you from Spotify. It's based on relatively unadaptable algorithm based on your long-term music habits and preferences. It reacts quite oddly to your immediate taste and situations and sometimes offers weird music choices you'd never come across. You'd need something more predictive and immediate mood and situation-based. How could GPT3 change this?
GPT-3 could fundamentally change the way we make, share, and listen to music. GPT-3 could be used to create a more personalized music listening experience by analyzing user input in real-time to suggest songs that perfectly fit the user’s current mood or situation. GPT-3 could also be used to generate music that is tailored to a user’s individual taste. Furthermore, GPT-3 could be used to generate lyrics and music based on input from a user. In this way, GPT-3 could revolutionize the way we make, share, and listen to music.
Google recently published research on MusicLM, a system that creates music in any genre with a text description. This isn't the first AI music generator. As TechCrunch notes, projects like Google's AudioML and OpenAI's Jukebox have tackled the subject. However, MusicLM's model and vast training database (280,000 hours of music) help it produce music with surprising variety and depth. The AI can not only combine genres and instruments, but write tracks using abstract concepts that are normally difficult for computers to grasp. If you want a hybrid of dance music and reggaeton with a "spacey, otherworldly" tune that evokes a "sense of wonder and awe," MusicLM can make it happen. The technology can even craft melodies based on humming, whistling or the description of a painting. A story mode can stitch several descriptions together to produce a DJ set or soundtrack.
MusicLM has its problems, as with many AI generators. Some compositions sound strange, and vocals tend to be incomprehensible. And while the performances themselves are better than you'd expect, they can be repetitive in ways human works might not. Don't expect an EDM-style drop or the verse-chorus-verse pattern of a typical song.
How could GPT-3 improve this? GPT-3 could be used to improve MusicLM in several ways. First, GPT-3 could be used to generate more personalized music based on user input. For example, GPT-3 could generate music based on a user’s favorite genre or artist, or even based on their current mood or situation. Additionally, GPT-3 could be used to generate more varied and complex musical compositions. GPT-3 could also be used to generate lyrics to accompany the generated music, allowing for a more immersive listening experience. Finally, GPT-3 could be used to generate music faster and with fewer resources, allowing MusicLM to scale up its operations and produce music more quickly.
As with other Google AI generators, the researchers aren't releasing MusicLM to the public over copyright concerns. Roughly one percent of the music produced at the time of publication was copied directly from the training songs. While questions regarding licensing for AI music haven't been settled, a 2021 whitepaper from Eric Sunray (now working for the Music Publishers Association) suggested that there's enough "coherent" traces of the original sounds that AI music can violate reproduction rights. You may have to get clearances to release AI-created songs, much like musicians who rely on samples.
AI already has a place in music. Artists like Holly Herndon and Arca have used algorithms to produce albums and museum soundtracks. However, those are either collaborative (as with Herndon) or intentionally unpredictable (like Arca's). MusicLM may not be ready for prime time, but it hints at a future where AI could play a larger role in the studio.
In the near future, GPT-3 could be used to generate music faster and with fewer resources, allowing MusicLM to scale up its operations and produce music more quickly. Finally, GPT-3 could also be used to generate music that includes original elements and variations, which could help reduce the risk of infringing on existing copyright laws.