MPEG-7 Music Player
By 2005, personal music devices will hold 10,000 or more songs. Such massive storage requires new interfaces for the user to be able to easily access content.
Background & Objective: Personal audio devices, such as MP3 music players, are extremely popular amongst music fans, especially those who like to download their content over the internet. Users can currently put up to 2000 songs on a single portable playback device, and will have access to up to 10,000 songs in the near future. The problem with these devices is that they lack any knowledge about the content and therefore offer no way for the user to navigate the enormous amount of information. MusicFinder addresses these problems by offering a query-by-example feature that finds lists of songs that are similar in style and mood to the current song.
Technical Discussion: MusicFinder is built upon MERL’s contributions to the audio part of the MPEG-7 international standard for multimedia content description. Songs are indexed using a hidden Markov model (HMM) trained on many different types of music, much like a speech recognizer, but for musical signals. Histograms of the HMM output sequences are used as features for similarity matching on the musical signals. Musical structure (intro / verse / chorus) is extracted by clustering on the within-song similarity matrix providing audio summary information. Style or mood ordered playlists are generated by sorting the similarity scores between all pairs of songs in the database. MPEG-7 defines the methods and structures for extracting the indexes from audio content.
Outside Collaborations: Song and artist metadata indexing was developed in collaboration with the Massachusetts Institute of Technology (MIT).
Contact: Anthony Vetro
Technology Area: Audio Video Processing
Modification Date: November 1, 2007
