Petronia Radio
Jan. 30th, 2009 04:14 pm- 17:22 curse of dimensionality: it's impossible to sample a 3-minute CD-quality stereo recording (2 to the 254,016,000 bit sequence) #
- 17:24 extract high-level features: MFCC (timbre coefficients), autocorrelation (temporal structure), spectrogram (pitch, timbre) #
- 17:26 aggregate features: 3-5 seconds is the most useful segmentation size #
- 17:31 GAMME (AdaBoost), Mandel & Ellis (SVM) - MIREX Artist Identification, Genre Prediction contests (MAGNATUNE and USPOP databases) #
- 17:32 currently looking at the segmented chart of this waveform (blue bar represents saxophone) ♫ blip.fm/~1vgr6 #
- 17:33 instead of predefined genres, data mine social tags from last.fm (...I'm realizing that tags on last.fm have LASTING SIGNIFICANCE) #
- 17:36 covariance of social tags - training set - reliably labelled artists - lack of negative tagging #
- 17:37 New Order is 0.954 correlated to "britpop" and 1.214 correlated to "electronic" (-1.475 correlated to "jazz") #
- 17:39 obvious weakness: most tagging is done at the artist level, but artists work in different styles (album- and track-level tags work better) #
- 17:44 most easily predicted tags: world, electronica, male lead vocals, "seen live" (apparently this is a virtual synonym for "indie") #
- 17:45 least easily predicted tags: comfortable, loving #
- 17:46 you can feed it a few seconds of the bassline from "Seven Nation Army" and it'll find stuff with the same timbre #
- 17:47 or Gabriel and Rodriguez's rhythmic features (flamenco guitar) #
- 17:50 isomap - shortest path through 2D tag space (descriptive tags vs artists) - from Mozart to Nirvana #
- 17:52 as dude points out this could make a neat iPod app with cover flow #
- 17:53 timbre-related tags: anarcho-punk, celtic, saxophone, female vocalists #
- 17:53 rhythm-related tags: eurodance, psytrance, minimal techno, video game music #
- 17:55 I see from this screencap that Muse is prominently tagged as "overrated" on the tastekeeper.com beta #
- 17:56 generative music for video games: analyze what's happening in the game, generate real-time soundtrack from player's music collection #
- 17:57 another possible use: blog recommender #
- 21:00 one of many non-considerations when Blur chose their band name: data mining via Flickr user tags #
- 01:04 top 20 artists tagged as "genius" on Tastekeeper: The Beatles, Bob Dylan, Bill Hicks (comedian), Tom Waits, Radiohead, Björk, John Lennon #
- 01:06 Miles Davis, Charles Mingus, David Bowie, Frank Zappa, Jimi Hendrix, Damon Albarn, Pink Floyd, Wolfgang Amadeus Mozart, Ludwig van Beethoven #
- 01:15 John Coltrane, Beck, John Frusciante, Johann Sebastian Bach #
- 01:20 top 20 artists tagged as "overrated": Radiohead, Beyoncé, Coldplay, Nirvana, Christina Aguilera, The Beatles, Metallica, Green Day #
- 01:21 Red Hot Chili Peppers, U2, Fall Out Boy, Linkin Park, Paris Hilton, System of a Down, My Chemical Romance, Muse, Slipknot #
- 01:24 Death Cab for Cutie, The White Stripes, Tokio Hotel #