Cocktail-Party Effect with 
Computational Auditory Scene Analysis
- Preliminary Report -

Hiroshi G. Okuno, Tomohiro Nakatani, and Takeshi Kawabata
(NTT Basic Research Laboratories)

One of important and interesting phenomena in sophisticated
human communications is the {\it cocktail party effect}:
that even at a crowded party, one can attend one
conversation and then switch to another one.  To model it in
a computer implementation, we need a mechanism for
understanding general sounds, and Computational Auditory
Scene Analysis (CASA) is a novel framework for manipulating
sounds.  We use it to model the cocktail party effect as
follows: sound streams are first extracted from a mixture of
sounds, and then some sound stream is selected by focusing
attention on it.  Because sound stream segregation is an
essential primary processing for the cocktail party effect,
in this paper, we present a multi-agent approach for sound
stream segregation.  The resulting system can segregate a
man's voice stream, a woman's voice stream, and a noise
stream from a mixture of these sounds.