Re: Blind Source Separation by Sparse Decomposition ("Richard J. Fabbri" )

Subject: Re: Blind Source Separation by Sparse Decomposition
From:    "Richard J. Fabbri"  <fabbri(at)NETAXIS.COM>
Date:    Wed, 8 Sep 1999 18:33:45 -0400

Al,

If you have ever heard (voltage comparator per my email) clipped
speech (nth approximation to Licklider's "infinitely-clipped" speech)
then, you know the remarkable job our hearing does perceiving the
spoken word via zero-crossings.  Yes, assumptions such as band
limiting and zero-mean reduce the universe of possible signal
reconstructions once given a measured set of zero-crossings.
But, these ideas can also be appreciated as schema satisfying
Time-Domain boundary conditions where the "boundaries" are
the zero-crossings.

My reason for proposing Binaural audition w/ clipping is probably
becoming quite clear but, I will not write that reason hear.  Anyone
who wants to witness a startling demo of "stream segregation",
otherwise known to Cherry as "The Cocktail Party Effect", should
simply playback a Binaural recording thru voltage comparators
and listen to those comparators through stereo headphones.

While listening to these reconstructed (Binaural) spatial sources,
recognize that we're not discussing theory ... we are:
        - Demonstrating psychoacoustics
        - Demonstrating source "segregation" via zero-crossings.

Primarily period & interaural delay are encoded by zero-crossings.
Most of the additional (directional) encoding, caused by pinnae and
head diffraction, have been removed from the Binaural signals found
at the L/R comparator outputs, i.e., some degradation of Front/Back
and Up/Down discrimination will be experienced while listening to
clipped Binaural recordings.

But, can anyone explain these reconstructed spatial sources via a
frequency analysis of "infinitely-clipped" (Binaural) channels?
If not, then, how SHOULD we approach the CASA problem?
Are there other tools and models we have ignored?

Finally, there are better approaches to zero-crossing reconstruction,
i.e., perceived with high fidelity but, these require a time-domain
theory that will make no general progress thru our community
while frequency analysis remains unquestioned.

Try the above experiment.
How do zero-crossings help us separate spatial sources?
Can you explain this zero-crossing effect using frequency analysis?


Rich