Classifying music into different genres is easy for humans and much harder for computers, but a new approach could yield a significant breakthrough in machine listening.
Most attempts so far have focused on getting computers to quantify parameters like tempo, loudness, melody and rhythm and matching those up to a list of manually sorted songs. But the researchers behind this new method believe their system is far simpler.
It classifies music by breaking it down into 88 frequency bands (like the different parts of an equaliser), computing the average power over several short-duration 'frames', giving a metric for pitch. For seven different music genres that the researchers tested, this metric was very distinct.
To narrow things down further, the tempo of the song, its variation in volume and its 'periodicity' (how often a song repeats) are thrown into the mix. With these four variables, the system was able to quickly and simply define a genre of a song.
When tested against earlier music classification systems, the researchers claim their results were "substantially better" - though it'll likely take some time to filter through into the systems used by music services like Spotify and the just-launched Apple Music. In the meantime, the researchers reckon it'd prove useful in digital archiving of music.
The details of the method were published in the International Journal of Computational Intelligence Studies.
Image credit: Tomas Fano // CC BY-SA 2.0
Article continues below