REAL- TIME SPEECH REC O GNITION SYS TEMS 
Hy Murveit 
SRI International 
OBJECTIVES 
SRI is developing the hardware, software, and Mgorithms necessary to achieve real-time speech- 
recognition and spoken-language systems. As the first phase of this project, SRI is currently 
completing the design of a prototype that will be able to recognize from 3,000 to 9,000 words of 
continuous speech using bigram language models. This prototype is based on four special- 
purpose integrated circuits. SRI is also continuing to improve the SRI DECIPHER speech recog- 
nition system, and enabling it to be incorporated into SRI's architecture for the integration of 
speech recognition and natural language processing. 
RECENT ACCOMPLISHMENTS 
• Specified and designed (in cooperation with U.C. Berkeley) four speciM-purpose integrated cir- 
cuits that will serve as the basis of our real-time large-vocabulary speech-recognition system. 
• Designed computer cards for the implementation of the prototype hardware. 
• Implemented software to run the the DECIPHER signal processing routines in real time on a 
TMS32025-bused board designed at U.C. Berkeley. 
• Improved the design of the DECIPHER speech recognition system, which achieves high accu- 
racy speaker independent or dependent continuous speech recognition. This included developing 
techniques for automatically generating pronunciation networks from standard training corpora. 
These networks improved speaker-dependent and speaker-independent speech recognition accu- 
racy. 
• Developed search strategies for DECIPHER that allow it to be efficiently used in SRI's 
spoken-language system framework. 
PLANS 
* Complete the prototype system and build three working real-time recognition systems. These 
systems will be evaluated with an application that uses a statistical language model. 
• Develop concrete specifications for real-time spoken-language systems. This will likely include 
modifications to the prototype system to allow it to work in SRI's spoken language system 
framework--primarily by adding hypothesis pruning to the prototype. 
• Continue improving the DECIPHER system including the addition of consistency modeling to 
the system, and real-time implementations of noise-processing algorithms developed at SRI. 
.458 
