New search engine helps you find new music

The technology behind a "similarity search engine" comes from research into what are termed Artificial Art Critics (AACs). AACs are composed of two components: an evaluator, which weighs how a typical human population would judge the aesthetic quality of a piece, and a feature extractor that identifies its general qualities. It's doubtful that the evaluator would be required for a search engine, as it might interfere with those searching for music that's generally considered aesthetically displeasing. That sort of music will probably pose a challenge to this technology in general, as the researchers note that it's "much harder to find truly unpopular (bad) music, since, by definition, the latter does not get publicized or archived."

That leaves us with feature extraction, which appears to be a three-part process. Previous work by the grant's author and others has produced neural networks that are trained to recognize both composer and style for a range of music including different eras of classical, country, rock, and jazz. Composer identification on classical pieces has a success rate of over 90 percent, while style attribution varied from 70 to 90 percent in a broader sampling of music. Presumably, any search engine based on this technology could verify its predictions against any metadata associated with the file and use that information to improve its success rate.