Re: Computerized genre analysis (Emmanuel Vincent )


Subject: Re: Computerized genre analysis
From:    Emmanuel Vincent  <emmanuel.vincent@xxxxxxxx>
Date:    Sat, 10 Nov 2007 18:57:59 +0100
List-Archive:<http://lists.mcgill.ca/scripts/wa.exe?LIST=AUDITORY>

Hi Larry, Much work has been done on this issue in the last 5 years. Existing computer systems are indeed trained on large databases and performance improves with the size of the database. You can have a look here first: http://www.music-ir.org/mirex2007/index.php/Audio_Genre_Classification_Results (click on the entrants' names to read their papers) A few studies have been done with human subjects with no prior training. The conclusion was that humans achieved similar performance to computer systems, say around 75% accuracy for 10 genres. http://ieeexplore.ieee.org/iel5/9248/29346/01326806.pdf However the notion of genre typically varies among people. So humans could probably achieve higher accuracy if they were subject to the same training procedure than computer systems. This has not yet been studied to my knowledge (DAn knows better than me though). Best, Emmanuel > Hi Folks, > Can anyone suggest research that has resulted in genre classification by computerized analysis? Humans seem to do this by making statements like "oh > that sounds a little like Bob Marley." The underlying similarities are clearly being analyzed against the listener's memories of remembered/heard > works. How would or has a computer gone about the same task? > I can imagine that a good algorithm would get better and better results with > more and more musical samples. True?? > Larry > Lawrence Borden > Associate Professor of Trombone, > Vanderbilt University > Blair School of Music > Principal Trombone, > Nashville Symphony Orchestra


This message came from the mail archive
http://www.auditory.org/postings/2007/
maintained by:
DAn Ellis <dpwe@ee.columbia.edu>
Electrical Engineering Dept., Columbia University