A Robust Preprocessor for Speech-Recognition Systems 
George Zweig, Principal Investigator 
Hearing Research Laboratory 
Signition, Inc. 
P.O. Box 1020 
Los Alamos, NM 87544-1020 
PROJECT GOALS 
Signifion is designing a preprocessor chip for speech- 
recognition systems operating in noisy environments. 
Current preprocessors, such as those based on moving 
Fourier transforms or linear predictive coding, are linear 
and not effective when speech is embedded in high lev- 
els of noise. 
Specifically we are: 
• Developing a nonlinear transmission line speech 
preprocessor that accurately simulates the mechanics of 
the inner ear over the entire range of sound pressure lev- 
els found in speech. 
• Preprocessing speech with the nonlinear trans- 
mission line and demonstrating that there is a substantial 
improvement in the signal-to-noise ratio at the output 
taps of the line as compared with the output of conven- 
tional linear speech preprocessors. 
• Assessing the feasibility of creating a digital 
transmission line on a chip to use as a preprocessor in 
the CMU, BBN, or MIT DARPA funded speech recog- 
nition systems. 
laser" that traps and amplifies quiet sounds, thereby aid- 
ing in their detection and analysis. This system was 
modeled as nonlinear transmission line with active cir- 
cuit elements and negative feedback. 
• Developed a fourth-order Runge-Kutta method 
for solving the nonlinear transmission-line equation. 
• Solved the nonlinear transmission-line equation 
for a variety of simple nonlinearities using clicks and 
pure tones as inputs. 
• Determined the form of the nonlinearity by com- 
paring solutions of the transmission-line equation with 
Rhode's corresponding data. 
• Showed that two frequencies interact with one 
another by affecting each other's region of amplifica- 
tion. Such two-tone suppression, the auditory analogue 
of visual lateral inhibition, leads to formant sharpening. 
• Developed a transmission-line parallel-process- 
ing algorithm for the Connection Machine. 
• Studied the feasilibility of creating digital VLSI 
chip preprocessors that emulate the nonlinear transmis- 
sion line. Preliminary results are encouraging. 
RECENT RESULTS 
• Deduced the mechanics of hearing at low sound 
pressure levels from measurements of basilar membrane 
motion. The mechanical properties of individual sec- 
tions of the cochlea have been determined by examining 
their collective response to tones of different frequency. 
The conclusion is that each section acts like a harmonic 
oscillator with negative damping. The oscillator is con- 
trolled by negative feedback that drives the oscillator 
with a force proportional to the displacement of the os- 
cillator at an earlier time. The time delay is approxi- 
mately 1.75 times the oscillator's period. Thus, the inner 
ear is "active," creating sound that interferes "intelli- 
gently" with the incoming sound. 
• Showed that the inner ear acts as a "subthreshold 
PLANS FOR THE COMING YEAR 
• Solve the nonlinear transmission line equations 
for noisy speech and provide a preliminary qualitative 
assessment of the signal processing capabilities of the 
nonlinear transmission line. 
422 
