Recently, there has been a revolution in the way that music has been delivered to users. The universal availability of broadband to the home and the development of cheap, high-capacity MP3 players has led to an exponential growth in music distribution over the internet, and to the emergence of large personal collections of songs held on users computers and players. This in turn has led to a need for effective techniques for organising, browsing and visualising music collections and generating playlists. Although metadata giving details of e.g. the track title, the album, the artists etc. is available for much of the music available on the web, it is not universal, and this data is usually not detailed enough to implement the above techniques to a high standard. We have been investigating techniques for automatically classifying the genre of a song and measuring the similarity of two songs using only the audio signal. I will describe our approach to these two related tasks, and present results that suggest it is possible to perform them with reasonable accuracy. I will also demonstrate our musical similarity software that suggests songs similar to an input song from a 5000 song collection.

Watch the webcast replay >>