Automatic music similarity retrieval aims to have computers find songs that are similar to other songs. Most successful similarity retrieval methods rely on human-annotated tags or social techniques. Content-based retrieval, on the other hand, attempts to design algorithms that allow computers to identify similarity based on the actual song content, i.e. the digital signature. As you might imagine, quantifying the aesthetics of a song is a difficult task, but it has the great advantage of not having to rely on meta-knowledge such as artist or genre about musical pieces.
In Bill Manaris’ lab at the College of Charleston, he and his students were engaged in this kind of research. The tech website Ars Technica featured an article on our work, though that link sadly no longer exists.
While on the topic, I’ll also mention some other related work we did in this group. Using quantitative features similar to those used in the similarity retrieval approach, combined with Genetic Programming and Artificial Neural Networks (ANNs), we developed an automated music composition system called NEvMusE (Neuro-Evolutionary Music Environment). Thanks to Bill Manaris, who kept track of these and put these up, you can listen to original compositions (one, by me :) based on evolved music pieces here: NEvMusE.