The weekend I took another look at the MusicMagic mixer by Predixis. My first attempt to use this system failed (apparently the 64-bit linux is not on the top of their priority platforms), but I do have a Mac Mini, and they support that so I tried it again on this platform. This time with much better results.

Using MusicMagic mixer is very easy to use. You point it at your music collection and it will go off and fingerprint and analyze the music. The fingerprinting process (similar to what MusicBrainz does) takes about 3 to 5 seconds per song. The fingerprint is used to query the Predixis database to see if an analysis for the song has already been performed, if so, the analysis portion can be skipped. If no analysis is found, then the song will be analysed and a number of musical features are extracted. The analysis is a very CPU intensive process and typically takes about 80% of the playing time. I pointed MusicMagic mixer at my collection of music (about 4,500 songs). Unfortunately, the fingerprinter did not findi many matches so most of my music will need to be analyzed. For this modest sized collection, this will take nearly 2 weeks of continuous CPU time on my poor little Mac Mini.

Fortunately, I don't have to wait until all songs are analyzed before I can play with the system. I can ask MusicMagic mixer to create an' 'instant mix' and it will generate a playlist of 10 songs that are acoustically similar. It does this by selecting a song at random from my 1500 analyzed songs and then selecting the 9 songs that are nearest to the seed song in this music similarity space. If I like the playlist generated, I can toss it into iTunes and listen to it, or I can easily generate another one. Once the initial analysis is performed, MusicMagic mixer is very fast, the playlists were generated instantly. The generated playlists, so far, seem to be quite good (but it is hard to tell with only 1500 songs analyzed). The songs in a playlist are typically in the same genre, and often include several songs by the same artist, but also include a variety of artists. I occasionally will encounter a 'clunker', a song that just doesn't belong (a harpsichord piece after a delta blues song). I'm looking forward to using the system once the full analysis is complete to see how well it performs.

The Predixis folks seem to know what they are doing and they seem to be hooking up with all the right partners. It will be interesting to see if they can make a go of this. The music recommender space is starting to attract a number of players, big and small, Predixis is the only one that is using automatic music analysis to drive recommendations, so they've carved out an interesting niche.


Post a Comment:

This blog copyright 2010 by plamere