Caltech Distinguished Graduate Student Lecture
		      (Everhart Lecture Series)
		    Monday May 18, 1998, 22 Gates
	    4:45pm Reception; 5-6pm Lecture and Discussion

-------------------------------------------------------------------------

		    Machines that Talk and Listen:
	       Speech Processing by Humans and Machines

		 Sam Roweis, Hopfield Group, Caltech
 
-------------------------------------------------------------------------

Imagine reading the New York Times if you were blind. Or having a
conversation with someone who speaks a different language.  How about
sending a fax or accessing your email while driving?  Speaking is the
most natural method of communication between humans. It would be great
if our machines could learn to speak and listen.  For decades engineers
have promised us that this technology is just around the corner.  But
so far, truly conversant computers have remained merely a promise.

What would we want them to do? Ideally machines should be able to to
read aloud (convert text into an acoustic waveform), transcribe
conversations (convert an acoustic waveform into text) and recognize
the sex, identity, or even mood of a speaker (voiceprint analysis).
Unfortunately, even the best speech synthesizers sound like a talking
moose, speech recognizers get about half the words wrong and speaker
identification systems are slow and unreliable. In comparison, the
best humans read eloquently, transcribe flawlessly and recognize
people they know after a few words. In addition, when these tasks are
made more difficult by adding noise and distortions, machine
performance degrades dramatically while human performance remains
almost unaffected. We must be using a very different strategy than our
current machines to solve these problems. This persistent gap between
the abilities of computers and people motivates speech researchers to
search for new approaches and ideas. 

In this lecture I will give an overview of the speech technology
field.  I will show how far we have come by demonstrating today's
state of the art systems and explaining how they work.  Finally, I
will discuss the future of the field by outlining exciting areas for
new research and briefly describing my own work which focuses on
applying models of human speech production to develop a new
representation of speech information.