See also 'The Echo Nest' (http://the.echonest.com/).
Ok, so I did investigate their API a bit (it's extensive!). I do mostly
compare with MusicIP (which I've been running for years), as I'd like to
have a replacement for smart mixing my own collection. Tests were run on
both my dev machine (about 3.5k tracks, MacBook Pro, SSD) as well as my
main server (20k tracks, Atom based, 2.5" HD).
As The Echo Nest is engine behind most of the larger music services' smart
features, I did obviously experience some of the same behaviour as using
those. For the differences to eg. Last.fm and Pandora, see this blog by
one of the EN founders:
http://notes.variogr.am/post/37675885491/how-music-recommendation-works-and-doesnt-work
After playing with various implementations of smart playlists (MIP,
Pandora, Last.fm, MOG, iHR, Deezer...) I'm glad he shares some of my
concerns about them.
Pro TEN:
- little to no installation effort (assuming there's a plugin available -
stay tuned ;-))
- available on any server hardware, even under-powered NAS devices
- support for Artist ("give me music from artists similar to A") and song
("give me similar tracks") mixes
- faster than MIP, both scanning/uploading as mixing on my (ymmv)
- they have a vast array of well documented options to tweak mixes,
including BPM, mood, "more/less of this artist", hotttness (yes, hottt!)
etc.
- Big plus for TEN: it's available and maintained today, while MIP is
dangling on a string
Contra TEN:
- no SB implementation available - yet
- the free API access is limited. Only a number of users would be able to
use a community driven solution.
- TEN doesn't know all my music, MIP does. I've got spotty mixes even with
big names like Peter Gabriel.
- I personally prefer MIP's mixes. Exploring some of the more advanced
parameters might improve this situation.
And here's a bit more information about some of above points:
- TEN requires you to provide them with some information about your local
music collection (artist name & track titles). These stay stored on their
servers accessible through some key. If you have privacy concerns: don't
touch it. (But then we had to upload hashes identifiying our music to MIP
as well, didn't we?)
- TEN doesn't need to scan your music. If you've got good tags, it's good
enough to upload that minimum set of metadata plus some key to get things
started.
- Uploading metadata for about 20k tracks to TEN was much faster on my
Atom based server than importing the corresponding data from MIP (not to
mention the scan time MIP needs itself). Creating a 75 track playlist
seems quicker using TEN than MIP in my case.
- TEN can only mix tracks it knows about. While they claim to have
information about 30mio tracks, it won't cover your neighbour's recording
of the previous week. Or some more exotic recordings like eg. the German
versions Peter Gabriel did of his 3rd and 4th albums. These latter
wouldn't mix at all. MIP can handle these, as it scans your music, no
matter what it is. Yes, you can upload such files to TEN to have them
analyze them, but I doubt this is a reasonable approach.
- Mix quality: IMHO it's easier to have eg a "rock guitar" mix with MIP
than with TEN. But then TEN's API is very powerful, and I've only
scratched its surface. But it's the same feeling I had with many of the
music services: TEN seems to rely more on meta information like "if you
like artist A, then you might like B as well" or periods, whereas MIP is
based on their analysis of the music itself. Eg. I'm always surprised how
MIP would queue up live recordings if I did seed it with a live track. It
must be the audience clapping or something which triggers this.
- TEN has tons of options to tweak. It's great to be able to tweak, but
providing this in a good UI is difficult.
- The free TEN API access is limited: they only allow a certain number of
queries per minute (20, upgradeable to 120). FWICT these are shared among
all installations using the same API key, and they explicitly forbid using
multiple API keys to work around this issue. OTOH they do mention other
open source projects which were given less restrictive limits. The same
applies to a limitation of 1000 profiles and mix session_ids: if I
implemented a plugin and it became very popular, this might be another big
issue, as there's at least one profile per installation.
As you can imagine I have not come to these conclusions without some
coding. Now let's hope I'll find some more time over the Holidays to
tinker around with this...
--
Michael
_______________________________________________
discuss mailing list
[email protected]
http://lists.slimdevices.com/mailman/listinfo/discuss
