Dragon Systems, Inc. 
90 Bridge Street 
Newton, MA 02158 
Tel. (617) 965-5200, FAX (617) 527-0372 
ARPANET DRAGON@A.isi.edu 
Principal Investigators: Janet M. Baker and James K. Baker 
Objectives: Dragon Systems is developing and building high performance, computationally- 
efficient interactive speech workstations, to sup'port adaptive speech recognition of large 
vocabulary, natural language speech input in real-time. These systems are based on the results 
of the multi-knowledge source (MKS) algorithm architecture and multi-processor accelerator 
hardware studies previously undertaken by Dragon under DARPA auspices, and on Dragon's 
separately developed speech recognition systems (VoiceScribe, DragonWriter, and DragonDictate) 
and related work. 
A highly interactive user interface is being further extended to facilitate system ease-of-use as 
well as performance and throughput. The integration of a full application interface will enable 
system testing under realistic operational conditions. Extensive testiqg will be conducted with 
users of these capabilities and their associated user / application / communication interfaces, in 
a series of challenging interactive performance tasks. The overall objective is to provide 
advanced speech recognition capabilities for operating environments in which the user has a 
demanding real-time task. 
Accomplishments in past 2 years: 
1. Demonstrated technical feasibility for multi2processor hardware architecture to support 
continuous speech recognition algorithms in a workstation. 
2. Designed flexible multi-knowledge source algorithm architecture to accommodate disparate 
knowledge sources in a consistent framework. 
3. Extended "McNemar's Test" to achieve greater sensitivity in significance measures of 
comparative recognition tests. 
4. Collected discrete and continuous speech database for Resource Management task (991 
words) with 7 speakers (about 50 K words total) in a moderate noise environment. 
5. Conducted tests and evaluated results on discrete RM database with DragonWriter-1000 
speech recognition system (includes 1 peripheral board and software) running in a 386-based 
MS.DOS personal computer. 
Plans for next 1 + years: 
During the initial phase of the present 3-year program, Dragon's continuous speech recognition 
algorithms will be ported into a multi-processor hardware and MKS software architecture 
hosted in an 80n86 (n>2) workstation. Dragon's on-line adaptive acoustic and language 
modelling algorithms will also be ported onto this platform as well as user and application 
interfaces. 
The primary performance task selected for demonstration and test-bed purposes is "Interactive 
Transcription". A user's speech will be transcribed interactively into text in real-time, through 
an application interface into a word-processing program for on-line report generation and 
document creation. It will also incorporate appropriate error correction and editing capabilities. 
Testing and the further development of evaluation methodologies will proceed in parallel. 
143 
