Fwd: [AUDITORY] multidimensional scaling of timbre

[Dan Ellis <dpwe@xxxxxxxxxxxxxxx>]

Dear List -

Michael Hall asked me to repost his message because it was not legible
in the digest edition of the list.

Best,

 DAn.


---------- Forwarded message ----------
From: Michael D. Hall <hallmd@xxxxxxx>
Date: Tue, Oct 21, 2008 at 8:40 AM
Subject: Re: [AUDITORY] multidimensional scaling of timbre
To: AUDITORY@xxxxxxxxxxxxxxx


The reduced number of dimensions from MDS solutions with natural
stimuli should reflect reliance on a few salient dimensions at the
cost of the not-as-salient ones. A particular dimension should not
necessarily reveal itself in the solution if several other dimensions
have already been relied upon by listeners to differentiate the
stimuli in the set. Thus, even though the dimension is something that
could potentially distinguish instruments in the set, it may not be
sufficiently strong to contribute to a particular solution (i.e., for
a particular stimulus set).

Much of the "noise" from MDS solutions can come not only from stimulus
uncertainty (e.g., when using a larger number of stimuli, or a set
that shares a greater degree of shared acoustic variation across
stimuli), but also from how the task is implemented. Because of the
large number of trials required to represent every pair in the set, it
is common for researchers to greatly limit the number of repetitions
of each pair. It is not unusual to see an MDS study with only 2 or 3
repetitions of a given pair. This means that any 1 response can
greatly impact average ratings for a given stimulus pair, even if that
response followed a trial where the listener was momentarily not
attending as well to one or both of the stimuli.

Another concern with reliance on MDS solutions with natural tokens is
in the correlational procedure used to determine the dimensions.
Acoustic dimensions often covary, which could produce misleading
information about the critical dimensions of timbre. For example, in a
paper under review, Jim Beauchamp and I recently instead used
identification and discrimination tasks to reveal that listeners rely
on an understanding of spectral envelope shape (through a manipulation
of formant structure) rather than on brightness alone (presumably
indicated by spectral centroid variation, which was manipulated in our
study by filtering). Yet, it is common for researchers to talk about
listeners' reliance on brightness based on a strong correlation with
spectral centroid variation in an MDS solution, even though other,
earlier work has shown even stronger predictions of performance based
on spectral envelope shape.

What I am suggesting is a need to identify the relevant covariates. If
a particular dimension contributes to one solution, but not another,
that can tell us that they are either related/redundant, or
alternatively, differ in salience, or both. It is then relatively
straightforward to put any related dimensions in competition with each
other (e.g., under more tightly controlled stimulus conditions) to
determine which is more generally relied upon in assessing instrument
timbre.

Sincerely,
Michael Hall
Department of Psychology
James Madison University

Michael Hall
Department of Psychology
MSC 7704
James Madison University
Harrisonburg, VA 22807
office: (540) 568-7877
fax: (540) 568-3322