clarke@csri.toronto.edu (Jim Clarke) (07/04/88)
A.I. SEMINAR, Tuesday, July 5, 11:00 am - 12:30 pm, SF3207
(SF = Sandford Fleming Building, 10 King's College Road)
R. Watrous
Siemens Research and Technology Laboratories
University of Pennsylvania, CIS Dept.
"Design, Training and Analysis of Connectionist Networks
for Speech Recognition"
The design of connectionist networks for speech recognition is considered
in the light of the characteristics of the speech signal. A network model
is defined, called the temporal flow model, which consists of simple pro-
cessing elements interconnected by links which have an associated weight
and time delay. The model is also general with respect to patterns of con-
nectivity, and will accomodate feedback links, within and between layers.
The model can be viewed as an adaptive dynamic nonlinear discriminant.
The model can be trained to categorize speech input data by a method of
mean-squared error minimization using gradient methods of nonlinear optimi-
zation. The complete gradient is shown to be computable for networks of
general connectivity.
Methods for analyzing network solutions are described, in order to inter-
pret and improve model performance. These methods include link weight and
unit activation analysis, parameter correlation, residual error and sensi-
tivity analysis.
The adequacy of the model for acoustic phonetic speech recognition is
explored along major axes of the acoustic phonetic space as defined by the
physiological theory of phonetics (Peterson and Shoup, 1966). Preliminary
results are reported for optimized networks for several experiments for a
single male speaker, using 10 to 40 training utterances, tested on 60 to 90
repetitions of each test token. The phoneme discrimination accuracies range
from 97% to 99%.
--
Jim Clarke -- Dept. of Computer Science, Univ. of Toronto, Canada M5S 1A4
(416) 978-4058
BITNET,CSNET: clarke@csri.toronto.edu CDNNET: clarke@csri.toronto.cdn
UUCP: {allegra,cornell,decvax,linus,utzoo}!utcsri!clarke