19.num/JW.wu
.ls 2
.na
.LP
Parallel Speech Recognition for the CNS-1

Su-Lin Wu

(Professor J. Wawrzynek and Adjunct Professor N. Morgan*)

NSF Graduate Fellowship and (JSEP) F49620-93-C-0014

For most people, speech is a natural and efficient manner of exchanging
information. Speech recognition research seeks to create machines that
can receive spoken information and act appropriately. Our imagination
easily provides many applications for such technology. However, most
of these applications far surpass the current state of the art.
Progress has been made slowly toward these goals, but the training and
testing of new systems in speech recognition research is very compute
intensive, and success has been limited.

This project involves the transformation of existing sequential code
for speech recognition into a parallel program for the CNS-1, a
connectionist network supercomputer. The structure of the problem is
inherently parallel and thus this application is a good candidate for
implementation on a parallel computer such as the CNS-1 (see
Connectionist Network Supercomputer). This machine has the potential
to remove many of the difficulties in speech research.

The program has two basic parts: a connectionist neural network that
estimates phonetic probability from speech features; and the Viterbi
Algorithm, a form of dynamic programming that uses negative log
probabilities as costs. We plan to implement an initial version of the
speech recognition code that is vectorized, but executes on only one
node. In later stages, the code will be revised to operate on the full
CNS-1 machine.

This program will be a practical application for the CNS-1, and also
help verify the design of the machine. Some parameters that will be
explored include latency and bandwidth between nodes, latency and
bandwidth to local memory and input/output capabilities.

*Adjunct Professor, EECS; ICSI