SOFTWARE TECHNOLOGY TRANSFER FROM SLS 
David Talkin, Principal Investigator 
Entropic Research Laboratory 
600 Pennsylvania Ave. S.E., Suite 202 
Washington, D.C. 20003 
PROJECT GOALS 
This summary describes a new project that is 
scheduled to start in March, 1992. 
In this project, Entropic will serve as a technology 
transfer agent for the Speech and Language Sys- 
tems (SLS) program. Our ultimate goal is to 
make complete systems developed in the course 
of SLS research readily available to interested 
parties in government, education, and industry. 
An intermediate goal is to provide clean, well 
documented implementations of speech recogni- 
tion systems such as SPHINX or BYBLOS for 
experimentation and application outside of the 
SLS program. 
The goals of the initial project are to create a 
technology transfer mechanism, to test it by 
means of a useful but modest-scale application of 
the mechanism, and to create a plan for the trans- 
fer of a full-scale speech-recognition system. 
RECENT RESULTS 
The end of the contracting gauntlet is in sight. 
PLANS FOR THE COMING YEAR 
Our first goal is to provide a portable implementa- 
tion of an "acoustic feature" front end. This pro- 
gram will accept sampled data in the SPHERE 
format developed at NIST, and produce a 
sequence of acoustic feature vectors. While 
there will be a wide variety of possible features 
that can be selected, there will be options to 
produce the acoustic feature sets that are used 
(prior to vector quantization) in the current sys- 
tems at certain SLS sites. The tentative list of 
sites is: BBN, CMU, MIT-LCS, MIT-LL, and SRI. 
A second goal is to establish a software repository 
and distribution mechanism for the acoustic fea- 
ture front end. This will be accomplished by an 
electronic mail server that will distribute informa- 
tion about the available software, and by ftp 
access for actual software transfer. For recipients 
without network access, tape distribution will be 
available at a nominal charge. 
A third goal is to create a plan (with cost and time 
estimates) for taking a full speech recognition 
system from within the SLS community and mak- 
ing it available to outside parties in a documented 
and supported form. 
To summarize, there will be three products of the 
initial effort: 
• an acoustic feature front end 
• a software repository and distribution mecha- 
nism; 
• a plan for transfer of a complete recognition 
system 
The transfer of a complete system is a possible 
follow-on project. 
Other possible follow-on projects include the 
transfer of additional analysis modules (e.g., 
HMM, VQ, etc.). The target modules would be 
chosen to be useful in constructing a specific 
speech recognizer while at the same time being 
technically interesting and useful on a stand-alone 
basis. 
473 
