Session 4: System Implementation Strategies 
Roberto Bisiani, Chair 
School of Computer Science 
Carnegie Mellon University 
Pittsburgh, PA 15213 
This is both a summary of what happened at the session 
and an introduction to the papers. Although the opinions 
expressed here are as balanced as possible, they might 
reflect my bias, for which I apologize in advance. 
The session was targeted towards the issues that are 
raised by the implementation of complete, real-time sys- 
tems. It seems that the Speech Community has finally real- 
ized the importance of demonstrating fully usable real-time 
systems. Two definitions of real-time seemed to be accept- 
able to the workshop participants. A system is real time 
either if: 
• it can keep-up with continuous input or if 
• it returns a fully parsed sentence within 200ms from 
the end of the utterance. 
The issues in implementing usable systems are: 
• recognition speed; 
• development cost and time; 
• production cost. 
The papers in the session concentrated on the first two 
issues. It was very nice to see that two papers out of three 
provided non-trivial speed improvements mainly with 
software techniques. 
The BBN paper, the first to be presented, hinged on a 
number of algorithmic improvements. The most notables 
being the reduction of the number of transitions in a gram- 
mar by means of zero states and the use of a suboptimal 
but fast Forward-Backward Search. It is difficult to 
precisely evaluate the relative improvement figures shown 
in the paper because they are measured with respect to the 
speed of an initial system of unknown (programming) 
quality, Nevertheless, the algorithmic speed improvements 
were substantially larger than the improvements due to the 
hardware (a state-of-the-art i860 board). 
The second paper, presented by Dragon, hinged on a 
technique, called rapid match, that cuts the number of 
hypothesis during the search by limiting it to a subset of the 
possible words. This technique makes it possible to imple- 
ment useful and impressive recognition systems on garden- 
variety 386-based personal computers. As with the pre- 
vious paper, the advantages provided by better hardware 
were much less impressive than the advantages made pos- 
sible by clever algorithms. 
A completely different approach was presented by SRI 
and Berkeley. The paper describes a custom machine that 
implements a heavily pipelined Viterbi search. Custom 
chips and a large amount of memory make up the bulk of 
the machine. The performance, at least on paper, is about 
two orders of magnitude better than the performance of 
current general purpose systems. Although this gap might 
be reduced to one order of magnitude by the inlroducdon of 
new general purpose systems, the performance of this sys- 
tem is potentially very impressive. The audience had ques- 
tions on technology (MOSIS CMOS) and on availability 
(sometime towards the end of the year). 
At the end of the session the chairman gave a brief 
progress report on the PLUS system being developed by 
CMU. This system was not described at the workshop be- 
cause it has already been described in computer architec- 
ture papers. PLUS is a distributed-memory multiprocessor 
composed of mesh-connected nodes. Each node contains a 
Motorola 88000, static and dynamic memory, and circuitry 
to make the local memory visible to all the other processors 
as if it were local to them. Systems with 1 to 64 nodes are 
possible. Each system is connected to a supporting 
workstation through its SCSI bus, facilities for input/output 
of digitized speech are provided. Construction of PLUS is 
proceeding and a few nodes will be running by the end of 
the summer. 
This, I believe, was a very positive session. It showed us 
that it will soon be possible to implement in real time: 
1. small but non-trivial tasks on commercial hardware; 
2. complex tasks that require fast search on custom 
hardware; 
3. full complex tasks, including natural language 
processing and data base search on semi-custom 
hardware. 
All these solutions will cost no less than a medium-size 
workstation. Should we start worrying about how to use 
algorithmic improvements and technology to build much 
cheaper systems? 
73_ 
