Soundbite for Songbird
SoundBite for Songbird is an add-on to the Songbird media player that uses audio similarity measures to create automatic “sounds like” playlists. SoundBite for Songbird also enables map and network based interactive navigation of music libraries. Instead of constraining you to metadata lists of text and art like most standard media players, SoundBite for Songbird allows you to explore your music collection in a new context relating to what the songs actually sound like. You can listen to familiar songs in a new light. You might even find songs you didn’t realize you owned.
Songbird is an open source, cross-platform application that uses Mozilla technology to provide an extensible, fully customizable media player. SoundBite for Songbird is feature complete for Linux systems and is in the final stages of implementation for Mac and Windows. Songbird users interested in getting an early look should contact me for a beta version.
This screenshot shows the Songbird application with SoundBite’s two-dimensional Map View. The left column is Songbird’s default method of showing libraries, playlists, devices, and bookmarks. Along the top is a set of filters that allow the library to be filtered by metadeta such as artist, album, or genre. The column on the right contains SoundBite’s options and controls. The large space in the center is the map based media view. Here, each song is represented by a square marker. In this image, the marker colors relate to the genre of each track.
The layout of the songs is derived from timbral audio similarity features and an efficient method of multidimensional scaling. The mechanics of the algorithm are in my thesis, but the basic premise is that songs that sound like each other are located in relative proximity on the map. Songs that sound different are far away from each other.
The visualization is a fully interactive browser. It works like Google maps, but the map is your music collection. You can click and drag to navigate, and zoom controls are in the upper left corner. When you roll over a track marker, the track context pane appears. This pane shows artwork, track metadata, and playback controls. It also provides controls for rating tracks, generating automatic playlists, and adding the track to a target playlist. Track markers can be colored by genre, album, artist, or rating.
The Neighbor Network feature is activated when you click on a track marker. All of the tracks in the media view are processed to determine how similar they sound to the selected track, and the most similar tracks are highlighted in a network. The image above shows several highlighted Neighbor Networks in a media view with a relatively small number of tracks. Similar tracks are located close together, but there is no guarantee that the closest tracks according to the similarity measure will always be the closest tracks on the map. This is a consequence of the high dimensionality of the song similarity space. There is no perfect two dimensional mapping.
To this point, the images have all shown the Map View in SoundBite for Songbird. In this view, all tracks in the underlying playlist or library are displayed on the map unless they are removed by the metadata filters. A second view, the Network View, allows direct interactive navigation of the similarity network. When you select a track in your library and open the Network View, the visualization shows your selected track (the seed track) at the center of the network. The most similar sounding songs in your library expand out around the seed track, and the distance from the seed track to any other track relates directly to the audio similarity measure. Here, I used Sonic Youth’s “Teen Age Riot” to seed my network.
The track markers in this image are colored by album. Not surprisingly, the coloring shows that many of the most similar sounding tracks to “Teen Age Riot” are also on Daydream Nation. Let’s take a look at a track from a different artist. To the lower right of “Teen Age Riot” is “Ergo Space Pig” by Guided By Voices. Clicking on the “Ergo Space Pig” marker does two things. First, all of the tracks in the “Teen Age Riot” network adjust their position to show how similar they sound to the new seed track, “Ergo Space Pig.” At the same time, new tracks that sound similar to “Ergo Space Pig” expand out to grow the network.
As I continue to explore my music library’s similarity network, I can listen to songs, build playlists, and rate songs. This is what the network looks like after a couple more expansions:
In addition to the interactive visualizations, SoundBite for Songbird can create automatic playlists from a chosen seed track. Here is a an example of a SoundBite playlist for “Run For Your Life” by The Beatles:
Automatic playlists can be filtered to exclude songs from the seed album or the seed artist. You can also choose to limit the playlist to one song per artist to maintain variation. SoundBite playlists currently use only audio similarity, so it can be quite interesting to browse your collection for similar sounding songs without considering genre, era, popularity, or any other context. For users who desire more contextual continuity, a similarity measure that includes context features is on the upgrade list. The first priority, however, is pushing the Windows and Mac versions to release so that all Songbird users can take advantage of the features in SoundBite for Songbird.