From: malcolm@interval.com (Malcolm Slaney)
Date: Tue, 23 Aug 1994 10:56:46 -0800
Subject: Avery Wang at the Hearing Seminar
Message-Id: <9408231756.AA19820@interval.interval.com>

Sound separation "month" continues at the CCRMA Hearing Seminar with a talk
by Avery Wang on his PhD research.  Some of you might have seen his work
last Spring.  We hope this talk will be more focussed on sound separation
techniques, and less on the mechanics of getting Avery's degree finished.
All of you who had burning questions to ask Avery, but didn't want to
embarass him, can bring your questions to Thursday's seminar.

        Who:    Avery Wang
        Why:    Indepth discussion of sound separation techniques!!!!!
        What:   Sinusoidal Sound Separation
        When:   This Thursday August 25th at 10:30AM
        Where:  CCRMA Library (Top Floor of the Knoll)

Next week, Dan Ellis will talk about his plans and work on sound
separation.  Dan is a student at the MIT Media Lab.

See you Thursday for Avery's talk, and the following Thursday for Dan's
discussion.  Bring your ideas about sound separation and see

-- Malcolm

From: Avery Wang <avery@ccrma.Stanford.EDU>
To: malcolm@interval.com
Subject: Abstract
Reply-To: avery@ccrma.Stanford.EDU


      INSTANTANEOUS AND FREQUENCY-WARPED SIGNAL PROCESSING TECHNIQUES

                       FOR AUDITORY SOURCE SEPARATION


                               Abstract

At the top level, my work deals with the problem of separating sounds,
for example, voice or specific musical instruments, out of a mixture.
My approach uses what I call Frequency Warped Signal Processing, in
which sinusoidal signal components are adaptively demodulated in order
to separate their AM and FM bandwidth components.  The demodulation
process requires the ability to track the instantaneous frequencies
(IF) of components of the input signal.  I introduce the idea of using
a Frequency-Lock Loop to track the IF of an analytic signal.
Additionally, I introduce the idea of using a constrained harmonic net
of trackers in what I call a Harmonic-Locked Loop (HLL) to compute a
minimum-variance estimate of the fundamental frequency of a harmonic
audio signal.  The minimum-variance harmonic tracking property allows
multiple rapidly slewing harmonic signals to be followed individually.
Once the signal parameters have been extracted the harmonic signals
may be isolated from the mixture and resynthesized.  A software
application illustrating sound separation will be demonstrated.
Examples include high-quality voice separation of excerpts from "Il
Barbiere di Siviglia", and others.