Residue-driven architecture for 
Computational Auditory Scene Analysis

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

The Residue-Driven Architecture presented here is a model of
auditory stream segregation from input sounds.  A subsystem
to extract auditory streams by using some sound attributes
is called an agency and the design of each agency is based
on the residue-driven architecture.  This architecture
consists of three kinds of agents: an Event-Detector, a
Tracer-Generator, and Tracers.  The Event-Detector
calculates a residue by subtracting the predicted input from
the actual input.  When a residue exceeds a threshold value,
the Tracer Generator generates a \tracer that extracts an
auditory stream from the residue and returns a predicted
input of the next time frame to the Event-Detector.  This
approach improves the performance of segregation and the
resulting system can segregate a woman's voiced stream, a
man's voiced stream, and a noise stream from a mixture of
these sounds.  Binaural segregation is also designed by the
architecture.