The Echo Nest Makes Pandora Look Like a Transistor Radio

I don't know if it's just me, but Pandora is a pretty low bar for music recommendations. I paid for Pandora One for a few years, but pretty much stopped using it after a few months because its playlists would be incredibly incongruous. My Radiohead station became overrun with, for example, Katy Perry despite my efforts to "thumbs down" songs that didn't fit the Radiohead sound that I wanted.

Invariably, every time I've tried Pandora since, it plays one or two songs that match (the same ones every time, which gets really boring) before it meanders off into a forest constructed of oddities and WTF moments.

iTunes' Genius recommendations / playlists, though, have been incredibly good for me. Relatively obscure / nuanced selections generate amazingly well-balanced playlists with good variety without straying too far from the sound I want to listen to. For example, RJD2's "Final Frontier" is followed by:

* DJ Shadow - Keep Em Close * Blackalicious - Rhymes For The Deaf, Dumb, And Blind * Wax Tailor - I Don't Know

Pandora, on the other hand, comes up with:

* Diddy - We Gon' Make It * T.I - Doperman * Diverse - Wylin Out

I know Pandora's selecting from a huge library and iTunes Genius is selecting from my library, but my library is incredibly diverse and is in no way narrow enough to improve Genius' results. I get great playlists for everything from Coconut Records to Mayer Hawthorne to The Black Keys to Johnny Cash to the Heartless Bastards.

Maybe Pandora was good for it's time, but it simply hasn't improved enough over time. It's as if Netflix still used their original recommendation algorithm.

tl;dr: Saying Echo Nest is better than Pandora is not a very strong statement.

I don't really get it. So it's a very big database of music metadata, where apps can pull information from, but it doesn't include the actual music. So if Pandora is a transistor radio, then The Echo Nest is a big encyclopedia. And in my world transistor radios play more music than encyclopedias.

Here's the million song dataset: http://labrosa.ee.columbia.edu/millionsong/

Last night I downloaded the SQLite db 'track_metadata.db' and ran some simple queries like sorting for the hottest artists (Kanye West, Daft Punk, Black Eyed Peas), most prolific artists (Ike and Tina Turner: 208, Michael Jackson: 204, Johnny Cash: 201), higher average song-length by year (1975: 258s, 1976: 255s, 2001: 254s).

I wish the SQLite db contained more fields like beats, loudness, song_hotness etc. as rickrolled here: http://labrosa.ee.columbia.edu/millionsong/pages/example-tra.... I don't think everyone needs the full waveform-analysis data present in the 280GB download.

I'd rather download a 2GB metadata file for multi-variable analysis like the Netflix prize instead of the entire 280GB dataset just to get the additional fields.

My end-goal was to do some analysis of what makes an artist/song 'hot'. If anyone is interested, I could redo something like http://chir.ag/projects/preztags/ but with music data.

Their music recommendation database has more songs and attributes. But theres no service that streams all the music they have indexed like Pandora does.

Meh. I've seen recommendations from The Echo Nest and they are absolute shit. I don't know what their algorithms are but they sure do need to change them. I am a huge fan of LabRosa though, so maybe their partnership will help them out.

If you want to see recommendations from a company that truly kick ass, check out BMat. Their technology works insanely well.

I applied for an internship at Echo Nest and never got a reply. It's not a big deal but it does make life easier for job hunters if companies respond with a definite 'no'. Even a form letter would be fine.