[sci.electronics] Speech/Pattern recognition using motion detectors

farjamit@eecs.cs.pdx.edu (Tom Farjami) (04/11/91)

     Greetings :

     I got couple of emails  regarding  the  use  of  motion
detection  for  Speech  recognition.  The  question that was
asked  was,"  would  this  circuit  arrangement  be  speaker
independent ?". I have expanded on this aspect.

     I am a Graduate student in the EE  department  here  in
Portland  State University. Currently the title of my thesis
is "a biologically  motivated  model  for  motion  detection
implemented  in  VLSI".  I  have  gone  through one round of
fabrication of this IC  and  the  preliminary  results  seem
promising.

     Here is a description of the project.  I am working  on
design  and  fabrication of a CMOS motion detector IC that I
have classified as being fuzzy. It has 10 digital inputs and
an analog output. A strong voltage on the output ( near VDD)
means that majority of the inputs were  fired  in  the  pre-
ferred  direction, otherwise  output is silent.  There is no
feedbacks and therefor no neural-net like back props in this
circuit.  It  detects  motion in real time; as long as input
signal's frequency is within few hundred to few kilo  hertz.
Circuit  is biologically motivated and operates in subthres-
hold.

     During my graduate  studies,  I  have  been  trying  to
answer  the  following  question,"Basically what are some of
the applications of this circuit in real life ?". I  thought
of  waveform  detection  in  scopes, pattern recognition and
even speech recognition myself.  You  might  ask,"How  would
that  be?".  Consider the following. If inputs of the motion
detector came from a set of comparators such that all of the
comparators  were  comparing  the same analog input signal (
say a speech signal) against  constant  (programmable)  vol-
tages.  That  is  each  comparator has a different reference
voltage but takes in the same analog signal  as  other  com-
parators  do.  Now if the analog waveform matched the wright
pattern in the preferred direction, which is also  our  time
domain,  motion  detector  will  turn on and we have pattern
detection. Of-course for this I need more  that  10  inputs,
perhaps  hundreds  to  give  me  high resolution and a state
machine plus a A/D converter to program  my  reference  vol-
tages with; but this can be done.  The power of this circuit
arrangement is that it doesn't do an exact fitting match  of
the  input  waveform with the reference voltages. But rather
it does a shape fitting match. Even more power-full  is  the
fact  that notion of time is removed from the reference vol-
tages; that is input waveforms of different frequencies  and
phases  can be matched against the reference pattern as long
as input waveforms have the same shape and a frequency range
of few 100Hz to a few 10 KHz . I believe this characteristic
is important if we are doing speech recognition and we  want
speaker independent matching.

     I believe Motion Detector project is very  interesting;
an extension of it could be an even more interesting project
capable of performing a very  complex  task  at  a  striking
ease.

     All comments will be appreciated.



        Regards,

        Tom Farjami

        farjamit@eecs.ee.pdx.edu



                       April 10, 1991