From: malcolm@interval.com (Malcolm Slaney)
Date: Sat, 27 Aug 1994 14:10:56 -0800
Subject: Dan Ellis at Hearing Seminar on 9/1
Message-Id: <9408272110.AA12594@interval.interval.com>

Dan Ellis will be leading the discussion at this week's CCRMA Hearing
Seminar.  During the last month we've heard two similar approaches to the
sound separation problem, both based on grouping harmonically related
sounds.  This coming Thursday Dan Ellis (MIT Media Lab) will be presenting
his ideas for a sound separation system.  Just what can we do when the
spectral energy in a sound is corrupted by an interfering sound?  This week
Dan will tell all!!!

        Who:    Dan Ellis (MIT Media Lab)
        What:   Models for Psychoacoustic Grouping
        When:   Thursday September 1, 1994 at 10:30AM
        Where:  CCRMA Library (Top Floor of the Knoll)

I've also got a schedule for the rest of September and I'm starting to work
on October.  If you know of any wonderful work on acoustics, let me know so
we can talk about the Hearing Seminar.

        9/1     Dan Ellis (MIT Media Lab) - Models for Psychoacoustic Grouping
        9/8     Jan Chomyszyn (CCRMA) - Distance Perception
        9/15    No Seminar (ATR Perception Workshop)
        9/22    No Seminar (ICSLP)
        9/29    Shahwan/Duda (SJSU) - Onset Detector Model

See you at CCRMA!

-- Malcolm

Models of psychoacoustic grouping for sound understanding and separation
                Dan Ellis (MIT Media Lab)

I have been working on perceptually-motivated representations of sound.
These discrete elements allow the explicit application of grouping
principles such as harmonicity, common onset and proximity.  Such rules
organize the sound into regions of energy likely to have arisen from
independent sources. This structuring can be used both for source
separation (by reconstructing individually-identified sources) and 'sound
understanding' i.e. explaining or describing the sound as the result of
some real-world causes.

Several lessons have been learned from this work so far.  Limitations of
the constant-Q sinusoidal model (as well as the success of mixed models
such as Serra's) have led to a richer representational vocabulary.  The
problem of spectral collisions has proved so taxing that I propose a system
(both as a computer model and as a speculative theory of the auditory
system) that extracts only weak bounds from the incident energy and infers
the fine structure of colliding elements from their context.  This will
result in a system capable of experiencing the 'continuity illusion' and
the successful reconstruction of clarinets corrupted by dropping cans.