Re: Could you provide some insight on general audio similarity measure?

[Brian Gygi <bgygi@xxxxxxxxx>]

Wang Dong,

A few weeks ago I posted to the list a message about a study I had done on
similarity of environmental sounds (non-speech non-music familiar
sounds).  The MDS solution to the similarity ratings revealed that one of
the best predictor of similarity among the sounds was the degree of
harmonicity of the sounds.  A lesser predictor was the structure of the
envelope (amount of silence, periodicity).  I can send you a handout of
that study if you like.

However, identification is a different story.  My dissertation was all
about factors in the identification of environmental sounds (in fact, that
was the title) and it can be found at:

http://zia.aisri.indiana.edu/~bgygi/

While some of the factors that predicted identification of filtered sounds
were similar to those in similarity study, (harmonicity, temporal
structure) it was not the case that the sounds which were rated as similar
tended to be equally identifiable.  In fact the correlation was almost zero.

What we could not find were a defined set of acoustic factors that were
reliable predictors of identification for all the sounds. The problem is
that environmental sounds are such a diverse class that making general
statements about the class of sounds is risky.  It may be that we listen to
different 'types' of sounds, e.g. vocalizations, machine sounds, weather
sounds, impacts, in quite different ways.

If you have any more questions about this, I would be glad to answer them.

Brian