For
the potential telephone company applications of Toll
and Assistance ('0' button) servicing credit card,
collect, and third party calls, and other applications
to be selected, FGC proposes to integrate the requisite
elements into a Speech Recognition (S/R) Server product.
In a partnership agreement between principal providers,
FGC would develop performance estimates, customize
interfaces, assemble and test the platform, measure
and tune the results, and demonstrate the product.
The
system integration effort will work to derive the
maximum effectiveness from interconnected elements,
employing the expandable processor resources and scalability
returns of the FGC/MP-7. The S/R Server may be organized
physically to reside on multiples of the base machine
architecture (essentially a seven-mode parallel multiprocessor,
expandable in increments of eight microprocessors).
The Server may also be constructed to include dedicated
recognizers for high and dense traffic levels co-resident
with multi-tasked recognizers that support several
speech channels simultaneously.
Design
targets will include a maximum number of speech channels,
minimum response times, and a unit cost per channel
that is best-of-class. This is consistent with overall
objectives within the industry to reach new levels
of affordability and efficiency for applications that
otherwise would fail the tests of business case and
end user acceptance.
-
The
FGC/MP-7 high performance platform
-
Speech
recognition software
-
Call
management system and software
-
Telephone
central office experience
-
Systems
engineering and systems integration skills
Characteristics
The
FGC/MP-7 parallel processor is uniquely suitable for
processing and recognizing continuous speech because
of its use of a binary tree architecture. This has
been described in a number of government related reports.
The IBM PowerPC chip at each of the microprocessor
nodes provides high performance floating point calculations
needed in processing speech recognition algorithms.
Pipelining and overlapped instructions assure a fast
path to resolutions and matches within the architecture.
Proprietary bus controls allow internal data transfers
at machine speed. Propagation delays are minimal for
the MP-7 among this class of parallel structures.
The combined attributes offer speech processing solutions
within required real-time constraints
Systems
Integration Approach
Systems
Integration steps include:
-
Select
the development site.
-
Determine
requirements for interfacing the FGC/MP-7.
-
Determine
requirements to modify the Recognizer System (e.g.,
BBN HARK) to best take advantage of the FGC/MP-7
parallel processing architecture.
-
A
modified version of the Recognizer will be compiled
and tested on the FGC/MP-7.
-
Demonstrate
processing and recognizing multiple channels of
continuous speech in real-time.
-
Measure,
test, tune, and modify to reach target objectives.
Collaboration
with premier speech providers and products is one
of several targeted areas. The others include financial
modeling, searching large data bases, e.g. Internet
search engines, and other combined compute,and data
intensive problem types as well.
-
Agreement
on collaborators
-
Agreement
on project objectives and timetables
-
Preliminary
business case and tests of reasonableness
-
Specified
roles and responsibilities
-
Assignment
of project resources (staffing)
-
Name
Project Manager
Speaker
Independent Connected Speech Recognition