PROGRESS REPORT FOR DARPA SLS PROGRAM 
AT DRAGON SYSTEMS, INC. 1 
Janet Baker and Larry Gillick 
Dragon Systems, Inc. 
320 Nevada Street 
Newton, Massachusetts 02160 
DRAGON@A.ISI.EDU 
TEL: (617) 965-5200 
FAX: (617) 527-0372 
INTRODUCTION 
This report provides a brief summary of recent work and 
descnbes anticipated future work in continuous speech 
recognition research at Dragon Systems under the auspices of 
DARPA's Spoken Language Systems project. 
REVIEW OF CURRENT WORK 
Dragon Systems has continued to develop its continuous 
speech recognition capability on personal computers, with a 
strong focus on achieving real-time or near real-time 
performance in that environment We are now able to run 
recognition using vocabularies of up to 5000 words, although 
the larger vocabularies do slow down the recognition 
substantially; therefore more work still needs to be done to 
improve the computational efficiency of our algorithms. 
The recogn~er now has an interface which allows for on- 
line adaptation. This interface constitutes a beginning step in 
the development of a "DragonDictate"-like er~ correction 
facility fet continuous speech, one that would allow a user to 
dictate text, correcting en~s as he proceeds, with the system 
gradually improving its models by using the feedback from the 
user's error corrections. 
In recent months Dragon has turned its attention to Resource 
Management, and this has resulted in a new focus on the 
special characteristics of this particular task; of course, Dragon 
continues to place a slrong emphasis on improving overall 
recognition accuracy in a way that will benefit the general run 
of speech recognition appfications. In order to use the 
digitized speech supplied on CD-Rom the first step was to 
write software that would closely emulate the old signal 
processing done on our standard hardware, so that we would 
have a baseline perf~rnance assessment that could be used to 
evaluate the new, more computationally demanding signal 
processing algorithms that we plan to implement in the near 
future. We then went through several cycles of development 
on our adaptation and Iraining alg~-ithms in the course of 
gradually improving our perfefmance on the RM1 development 
test data and are still in a period of rapid development. To 
enhance our ability to perform experiments with the large 
quantity of speech data that has been supplied, we ported our 
recognizer to the IBM RS-6000 workstation. We have a 
collection of these machines and, now that they have been 
networked together, it is possible to do experiments on many 
speakers simulmeously. The recognizer has also been ported 
to the Apple Macintosh computer. 
With our focus on improving the accuracy of the recognizer 
has come a concern with the sources of error. To enhance our 
ability to study our sources of error, we have developed a 
diagnostic program known as ERRSPEC, which displays 
segmented spectrograms of utterances and models, together 
with a variety of revealing plots that highlight where the 
recognition algorithm has gone wrong. 
FUTURE PLANS 
Dragon plans to continue its work on the Resource 
Management Task and to begin work on ATIS in the coming 
months. A primary goal will continue to be improvements in 
our overall accuracy, but with the additional aim of moving to 
a mcfe broadly based speaker independent mode~g strategy, 
one that is likely to be based on mixture distributions (at the 
PEL level, at the PIC level, and at the word level). Investigation 
of alternative signal processing algorithms will also be a high 
priority, as we move away from the computational constraints 
that our old hardware had placed on us. Post processing 
slrategies based on the outputs from the N-Best algorithm will 
also be explored. 
1. "ntis work was sponsored by the Defense Advanced Research Projects Agency and was monitored by the Space and Naval 
Warfare Systems Command under Contract N000-39-86-C-0307. 
414 
