On Jun 22, 2006, at 12:36 PM, Malcolm Slaney wrote:
On Jun 21, 2006, at 2:54 AM, Olivier Tache wrote:
I have read a number of "classical" papers about MDS and auditory dissimilarity (by Gordon&Grey, Grey&Moorer, Wessel) (and was wondering if such experiments were still carried out).
I think the Gray/Wessel approach has failed..
Failed at what? Malcolm, I think you have missed the point. My original goal in the early 70's was to develop representations of musical material that would help me reason about the composition of timbre sequences known as klangfarbenmelodies. In one of my early papers ( http://xenakis.ircam.fr/articles/textes/Wessel78a/ ), I demonstrate that one can obtain interpretable geometric representations of musically useful timbres and that such "timbre spaces" can be used to make predictions about the behavior of timbre sequences. Hardly a failure! Grey and Gordon demonstrated how such MDS spaces can be used to make interesting timbral interpolations or hybrid instrument sounds. Recently, I've taken a new look at the timbre space representations obtained by myself, Grey, McAdams, Wedin and Goude and am struck by how well Les Atlas's Modulation Spectrum describes what is a common feature of many of the 2-D spaces wherein one of the dimensions is related to the spectral envelope and the other the temporal envelope.
I have no objection to your and Terasawa's approach of testing a pre-ordained model such as one based on MFCC's. However, such tests should be carried out in a direct manner as suggested by Krantz and Tversky in their work on the foundations of the geometric representation of perceptual data (see Suppes, Krantz, Luce, & Tversky's Foundations of Measurement Vol 2). I doubt that MFCC's will pass the straightforward qualitative test of "interdimensional additivity" essential to a geometric representation.
"The spectrum is not an interesting steady date." she said as I was enveloped.
it's too hard to figure out what the results mean. (Just trying to be blunt to get your attention. ;-) You start with convenient sounds, measure perception and then try to figure out what the MDS dimensions mean. That hasn't worked. I think that is why people have not been pushing on it very hard lately.
Hiroko Terasawa and I have been taking an opposite approach. We're *starting* with the dimensions, synthesizing sounds and then measuring the stress between human perception and the pre-ordained model. Several papers describing our initial results are online at
Sounds like Jim is doing something in between the two extremes.