Wednesday, May 18, 2011

Avishai Cohen, according to Pandora

Avishai Cohen's thoroughly enjoyable "Nu Nu" just came up on Pandora.  (Link here;  be advised that the sample is not really representative.)  Perhaps it was the snarkicist in me, but I just had to know why Pandora included the piece.  Here ya go:


Well, at least they included "odd meter" and "minor tonality" as quantifiable attributes, but even then, is "minor tonality" accurate when the primary scale employs a minor second and a major third?  And if "a twelve-eight time signature" is a viable attribute in another tune, how hard is it to add another one for, say, "seven-four"?  (Keeping "odd meter" is not bad, just add the specific tag too.)

The rest is kinda funny.  Apparently "acoustic piano accompaniment" was worth mentioning (and just a bit wrong;  the piano plays the main figures, while being decorated by the bass and whatever instrument prompted the "unique instrumentation" reference...I'd guess oud based on the sound, but could well be wrong), but not the instruments that give the piece its essential texture, mood and identity.  (The bass in this piece is simply magnificent, and the "oud" sound plays beautifully against it.  Their improvisations in and around the Hebrew scale just come alive.)

I've got no idea what "light drumming" is.  Do you?

"Based on what you've told us so far", there is absolutely no reason to recommend something to me based upon the presence of drums of any kind--or piano, for that matter.  My patterns (for this station, which I intend to be fairly narrow) for thumbs-upping and thumbs-downing tunes should easily reveal that the number of songs I've thumbed up that do have drums are countable on two hands (and almost all of those will be tabla pieces), and the number that I've thumbed down that have drums are well over 50% of the thumb-down total.  Same with piano.  (Think "jazz piano trios" and "blues and folk singers" here, which still seem to reliably follow anything by John Zorn/Avishai Cohen and Davy Graham, respectively.)

Incidentally, when one looks up the tags for "Nu Nu", one finds a 100% correspondence with the list presented above.  So:  it matches me with every attribute listed.  (Other tunes show a subset of the total for "based on what you've told us so far";  although these often seem to be more random than data-driven, it is not the case that all songs list all attributes.)

Look, Pandora:  I'd be happy--happy--to have a listing both of the things about the current song that do match things that I've thumbed-up reliably, and also a listing of the things in the song that are different from what I've thumbed-up.  That might be a good way to expand one's horizons.  But the system regularly suggests that I've established a pattern of wanting to hear protest folk, popular blues, piano, drums, and most astonishingly, vocals.  The degree of "fail" in that is impressive;  it suggests rather that you haven't been listening to "what I've told you" at all.

And yet, between these fits of inexplicable (and repeated) connecting of unrelated dots, I continue to get some really great music from the service--did I not find both Avishai Cohen and Davy Graham in the first place, from an origin point of Dan Crary?--so it's not like it's all wrong.

Please, try and improve the data-driven excludes.  My suggested start is to recognize that when someone reliably thwacks every single vocal tune that has ever come up in the station, he just might not want to listen to vocals.  (Hint:  it's not a "by artist" thing.  Some people play multiple kinds of music, singing on some and not singing on others.  If the attribute belongs to the song rather than the artist, you'll significantly reduce your false hits.)

No comments: