ROBUST CONTINUOUS SPEECH RECOGNITION 
TECHNOLOGY 
PROGRAM SUMMARY* 
Clifford J. Weinstein and Douglas B. Paul, Principal Investigators 
Lincoln Laboratory, M.I.T. 
Lexington, MA 02173-9108 
PROGRAM GOALS 
The major objective of this program is to develop and 
demonstrate robust, high performance continuous speech 
recognition (CSR) techniques focussed on applications in 
Spoken Language Systems (SLS). The effort focusses on de- 
veloping advanced acoustic modelling, efficient search tech- 
niques, rapid enrollment, and adaptation techniques for ro- 
bust large vocabulary CSR. An additional Lincoln goal is 
to define and develop application of robust CSR to military 
and civilian systems, and to expedite effective technology 
transfer. 
BACKGROUND 
The Lincoln program began with a focus on improving 
speaker stress robustness for the fighter aircraft environ- 
ment. A robust hidden Markov model (HMM) system was 
developed with very high performance under stress condi- 
tions. The robust HMM techniques were then extended to 
yield state-of-the-art performance on the DARPA Resource 
Management corpus, using a tied-mixture HMM CSR ap- 
proach. 
Recent work has focussed on the large-vocabulary Wall 
Street Journal (WSJ) corpus, with vocabularies of 5K, 20K, 
and up to 64K words. The HMM CSR has been converted 
to a stack-decoder-based control strategy to operate effi- 
ciently with good performance in these tasks. 
RECENT ACCOMPLISHMENTS 
Recent accomplishments include: (1) development of the 
stack decoder and demonstration of its effectiveness on vo- 
cabularies up to 64K words; (2) development and integra- 
tion of fast-match and detailed match; (3) further develop- 
ment of acoustic modelling techniques for the large vocabu- 
lary task; (4) a full set of evaluation tests in the November 
1992 WSJ tests, including (e.g.) a 4.5% error rate on a 
5K speaker-dependent test; (5) development of recognition- 
time speaker adaptation techniques with substantial im- 
provements due to adaptation from both speaker-specific 
and speaker-independent initial models; (6) participation 
in and contributions to development of the WSJ corpus, 
*THIS WORK WAS SPONSORED BY THE DEFENSE AD- 
VANCED RESEARCH PROJECTS AGENCY. THE VIEWS 
EXPRESSED ARE THOSE OF THE AUTHOR AND DO NOT 
REFLECT THE OFFICIAL POLICY OF POSITION OF THE 
U.S. GOVERNMENT. 
including providing baseline language models to all sites; 
(7) survey and study of opportunities for military and gov- 
ernment applications of spoken language technology, and 
organization of a workshop focussing on technology trans- 
fer; and (8) continuing leadership of the DARPA spoken 
Language Coordinating Committee. 
PLANS 
Plans for the current program include: (1) development of 
advanced acoustic modeliing techniques; (2) development 
and improvement of stack-decoder-based HMM for large 
vocabulary tasks, via development and integration of ad- 
vanced acoustic models, acoustic fast match, and efficient 
search techniques; (3) development of technique for integra- 
tion of stack-based CSR with natural language processors; 
(4) extension of run-time adaptation techniques to adapt 
acoustic parameters of the tied-mixture HMM to speaker 
channel, and environment; and (5) continued investigation 
of applications opportunities for spoken language systems. 
400 
