Individuals' music collections that are stored on a
computer, whether legal or not, are often quite large and consist of many
different artists. This unique research paper from Hewlett-Packard outlines a
system that can automatically identify the singer of a song based upon digital
samples of that song, allowing for hassle-free sorting of many titles. The
system is first trained by exposing it to a representative sample of songs from
a wide range of artists. Then, by extracting sound features from different songs
and analyzing them, the software can make a best guess as to which artist is
singing. The system is shown to have an accuracy of 80 percent when trained with
a single song for eight singers and tested on 45 other songs by the same
singers.