Using Machine Learning to Isolate and Clean Vocal Tracks
Harriet Rogers
2026
Bio
I am a third-year Physics with Astrophysics MPhys student intending to pursue a career in computational astrophysics and cosmology research. I have strong interests in music (specifically The Beatles!) and in applications of machine learning so used this project as an opportunity to combine these interests in a new way.
About the Project
Machine learning and artificial intelligence (AI) tools have been used to produce music for decades. However, this has only been drawn to public attention in recent years due to the prevalence of large language models and other AI tools, alongside significant media hype. A particularly prominent recent use of AI in music was by the Beatles to posthumously release John Lennon’s song, “Now and Then”. They used a machine learning algorithm to separate Lennon’s vocals from his piano track and remove background noise so that the vocaltrack could be combined with newly recorded instrumentals and vocals from the other bandmembers, and the song finally released.
In this report, I describe the process of developing a similar AI software using a neural network to separate and clean vocals from audio tracks, as well as a consideration of the legal and ethical issues that come with such software. The software I developed was not as effective as I had hoped, largely due to the limited time and computing resources I had to train the model. However, it serves as a good proof of concept that it is relatively easy to develop this type of software, leading to questions about the future of this technology.