Brian Whitman,one of the founders of the Echo Nest, gave a provocative talk last week at Music and Bits. Some excerpts:
Useless MIR Problems:
- Genre Identification – “Countless PhDs on this useless task. Trying to teach a computer a marketing construct”
Hard but interesting MIR Problems:
- Finding the saddest song in the world
- Predicting Pitchfork and All Music Guide ratings
- Predicting the gender of a listener based upon their music taste
On Recommendation:
- “The best music experience is still very manual… I am still reading about music, not using a recommender.”
- “If we only used collaborative filtering to discover music, the popular artists would eat the unknowns alive.”
- “The SQL Join is destroying music”
Brian’s notes on the talk are on his blog. The slides are online here. Highly recommended:
Music Machinery 
#1 by lfranchi on October 28, 2009 - 10:46 pm
slide 110 makes me excited
“we will soon make all of our acoustic data available for searching and browsing”
can’t wait to build some great features into a certain desktop media player :)
#2 by brian on October 28, 2009 - 11:50 pm
yes, it is very exciting!! There was even a more exciting part relating to that new feature that was cut from the talk at the last minute due to lack of willing perfomer
ps, anyone that does genre ID that wants to write me hate mail, please be aware that I have once done genre ID and I understand the academic interest in predicting a known label on perpendicular form of data. I think mnb will post my actual talk, which will make me sound less like a Strange Person with Ideas (maybe)
#3 by Zac on October 29, 2009 - 3:53 pm
Very fun presentation.
Interested to know where Brian got the stats about the “Known Artists” and “Cost to Understand One Album” on slide 72.
#4 by brian on October 29, 2009 - 4:02 pm
we of course don’t know for sure about AMG & MGP, but we just extrapolate from various pieces of data we found and hear over the years… The point i said aloud (barely) was that it just shows the three different approaches and why EN’s is somewhat more efficient.. i don’t think anyone can argue that the EN analyzer/crawlers are slower or more expensive than a MGP or AMG editor.
there should be a latent disclaimer across the whole talk that it is not an academic presentation. The EN data is correct to the best of my knowledge, however :)