GSoC proposal draft: Spatial collection with fuzzy queries

[Soren Harward]

On 3/2/09, Teo Mrnjavac <teo.mrnjavac at gmail.com> wrote:
> Thanks. Actually I was planning to use PCA to let the acoustic
> features position the tracks rather than a SOM which I don't know that
> much about but I guess either way would be fine as a start given the
> experimental nature of the project. The part I'm less acquainted with
> is the acquisition of the needed acoustic features to feed to the PCA,
> but I know that it's definitely doable as it has already been done in
> a closed-source Java research project, so that's the part I would most
> need help with.

As far as acoustic features go, I've primarily looked at programs
which have participated in the annual ISMIR contests.  One of them,
the MARSYAS software suite developed by George Tzanetakis, is the
basis of GTACFeat, which is the small acoustic fingerprinting library
I developed for use with Amarok (and other similar music programs).
There's plenty of published work about different approaches, and
plenty of open-source signal processing software available.  The trick
is just putting the two of them together (like I did with GTACFeat).

As far as the presentation goes, I suggest you strongly consider SOMs,
because they present information in a way that's more intuitive for
low-dimensional (eg, 2D) presentations than PCAs.  I admit that this
is an opinion, and I'd be happy to explain it off list, since it's
getting pretty far off topic.

-- 
Soren Harward