Oral-History:James Kaiser

About James Kaiser

James F. Kaiser

Dr. James Kaiser was born in 1929 in Piqua, Ohio. He attended the University of Cincinnati, where he earned his electrical engineering degree in 1952. He then entered graduate school at the Massachusetts Institute of Technology (MIT), from which he received his S.M. and Sc.D. in 1954 and 1959, respectively. Kaiser's early work in signal processing was at Bell Laboratories, which he joined in 1959. In the Bell Labs Research Department Dr. Kaiser worked on projects such as acoustic concentration with microphones, improving speech signal processing systems, developing filter design algorithms, and spectral window research. His focus was on filter design, particularly in those which would convert analog data to digital for various communications purposes. In the 1980s, he concentrated on nonlinear approaches to filter design. When the Bell System broke up in 1984, he moved to Bellcore, the newly established research operation owned jointly by the divested local Bell telephone companies. He became an IEEE Fellow in 1973 [fellow award for "contributions in digital signal processing and the synthesis of digital filters"], and has received many other IEEE honors and awards, including several from the IEEE Acoustics, Speech, and Signal Processing Society. These include the ASSP Technical Achievement Award (1977), Meritorious Service Award (1978), and Society Award (1981). He is a recipient of the IEEE's Centennial Medal, and the W.R.G. Baker Prize (1995). He has been very active and has held several offices within the IEEE, including the Editorial Board. He received the Bell Laboratories Distinguished Technical Staff Award in 1982.

The interview spans Kaiser's path breaking career, especially his digital signal processing (DSP) work with Bell Labs and Bellcore. Kaiser discusses his graduate work at MIT, his association with the LearJet Company, and his decision to join Bell Laboratories, before describing his experiences with Bell Labs and his contributions to the DSP field. He describes his work with acoustics signal processing, including his development of filter design algorithms, his efforts to improve speech signal processing systems, his work with the block diagram compiler (BLODI), and the uses of multiplexers. Kaiser also explains his work with computers in the DSP field, the difficulties with strictly linear DSP models, and the major areas in signal processing today. He outlines DSP uses for the musical and film-making fields, and the necessity of knowing the scientific basis of practical DSP applications. The interview concludes with Kaiser's discussion of his recent interest in image enhancement.

Other interviews detailing the emergence of the digital signal processing field include C. Sidney Burrus Oral History, James W. Cooley Oral History, Ben Gold Oral History, Wolfgang Mecklenbräuker Oral History, Russel Mersereau Oral History, Alan Oppenheim Oral History, Lawrence Rabiner Oral History, Charles Rader Oral History, Ron Schafer Oral History, Hans Wilhelm Schuessler Oral History, and Tom Parks Oral History.

About the Interview

JAMES KAISER: An Interview Conducted by Andrew Goldstein and Janet Abbate, Center for the History of Electrical Engineering,11 February 1997

Interview #323 for the Center for the History of Electrical Engineering, The Institute of Electrical and Electronics Engineers, Inc.

Copyright Statement

This manuscript is being made available for research purposes only. All literary rights in the manuscript, including the right to publish, are reserved to the IEEE History Center. No part of the manuscript may be quoted for publication without the written permission of the Director of IEEE History Center.

Request for permission to quote for publication should be addressed to the IEEE History Center Oral History Program, 39 Union Street, New Brunswick, NJ 08901-8538 USA. It should include identification of the specific passages to be quoted, anticipated use of the passages, and identification of the user.

It is recommended that this oral history be cited as follows:

James Kaiser, an oral history conducted in 1997 by Andrew Goldstein and Janet Abbate, IEEE History Center, New Brunswick, NJ, USA.

Interview

Interview: James Kaiser

Interviewers: Andrew Goldstein, Janet Abbate

Date: 11 February 1997

Place: New Brunswick, NJ

Undergraduate education, Univ. of Cincinnati

Goldstein:

Let's start with your education. You went to MIT as an undergraduate?

Kaiser:

No, as a graduate student.

Goldstein:

I see.

Kaiser:

My undergraduate work was done at the University of Cincinnati. I graduated from the University of Cincinnati on a Friday and I began at MIT the following Monday. I came in there as a graduate student about 9:30 in the morning and reported to the Graduate EE Office, where Miss Young was the secretary. When I came in, she looked at her watch and said, "You're late! The class is already meeting! You better get up there." I said "Wow!" That was my introduction to MIT and off I went to class. I essentially spent seven years at MIT.

Goldstein:

Can I step back to Cincinnati for a second? Was that a conventional engineering program?

Kaiser:

Right. It was an electrical engineering department. Of course you know that the University of Cincinnati was the pioneer cooperative school for engineering education, so you could not go through the engineering school without participating in their cooperative program. It was a five-year program alternating between school and work. Of those five total years you spent 104 weeks (two full years) working in industry. You literally went the entire year round. You didn't have any real time off. You had two one-week periods off and one three-week period off, for a total of five weeks off in the year. But it was a very good program that I really thoroughly enjoyed.

Goldstein:

Do you think that experience gave you a special ability with applied, practical, problem solving?

Kaiser:

Of course. The company that I co-oped with for nine of the eleven periods was Lear-Avia (which later became Lear-Siegler). At that time they were primarily manufacturers of electrical equipment for aircraft: small autopilots, actuators, electric drives, and all that. Actually, during the Second World War, Lear had a manufacturing plant in my hometown in Ohio, so I knew their equipment literally from top to bottom. As a high school senior I knew almost everything that Lear manufactured in our town just by looking at the junk that they threw out. That's where I got interested in electric motors and actuators. During my fifth year at the University of Cincinnati, I became very interested in automatic control. The main book available on the subject was volume 25 of the MIT Radiation Laboratory series, Theory of Servomechanisms, by James, Nichols, and Phillips, which had just come out. So I had my copy of that. We didn't have a course in control at the University of Cincinnati, but I read that book and got very interested in the subject. That was one of the reasons I applied to MIT, because they had a very active program in automatic control.

Goldstein:

When did you graduate from the University of Cincinnati?

Kaiser:

1952.

Graduate studies, MIT

Kaiser:

Right away I had that Research Assistantship at MIT in their Servomechanisms Laboratory, which later became the Electronic Systems Laboratory. The main professor I worked with there was George Newton. After being a research assistant for four years I became an Instructor in the EE department, which was a very good experience. As a graduate student, I was teaching the graduate controls course and, at the same time, the undergraduate controls course with my mentor George Newton. During the second year of my instructorship, George was writing a textbook on automatic control system design. He was so involved as associate director of the servo lab that he wasn't finding time to finish that textbook, so he brought Lenny Gould, who was an assistant professor at that time, and me in to write the last half of the book. So I wrote chapters 5 and 6, plus two appendices, Lenny wrote chapters 7 and 8 and two appendices, and George wrote the final chapter, 9, and thus the book got finished. That book was Analytical Design for Linear Feedback Controls by George C. Newton, L. A. Gould, and J. F. Kaiser. I don't think I fully appreciated the value of that effort at the time. It was a book that had a rather narrow message to it. It became a real classic, and that didn't do me any damage when I applied to Bell Labs for a research position.

Bell Labs, 1960s

Kaiser:

When I finished my Sc.D. degree, in February of 1959, I received a three-year appointment as an assistant professor at MIT. I had gotten to the point I felt I had been doing too much theoretical and analytical work. I decided to take a leave of absence for a year to go to Bell Labs and see what the real world is like. So I interviewed at Bell Labs and got that position. I went down to Bell Labs and enjoyed it so much that I extended my leave-of-absence from MIT for another year. It quickly became obvious I liked Bell Labs very much and was able to do the kind of things that I wanted to do there, so I just stayed.

Goldstein:

Were you in the research department?

Kaiser:

Yes I was. The research arm of Bell Laboratories is only about seven or eight percent of the Laboratories. And even that side of the Labs is much more applied than what I had been doing at MIT. The one gentleman at Bell Labs I worked with who had probably the biggest influence on my work and thinking was Hank McDonald. I worked with and for Hank, and he was a big supporter of my work. Hank had finished his Ph.D. at Johns Hopkins, I think, in the mid-50s—maybe '56—and then came to the laboratories. One experience I always remember with Hank was when I was doing some filter designs for speech work. He came up to me and said, "Look, Jim, I don't care how small or how large your performance indices are, if it doesn't sound right, it isn't any good." So there was the real world clashing into the theoretical world. In the theoretical world, you come up with numbers and expressions. But in the real world, if it doesn't sound right, it isn't any good, no matter how small you measure the performance errors to be. That was Hank.

Goldstein:

Let's talk about that later, this notion of instrumentalism versus a more subjective position with respect to results.

Kaiser:

<flashmp3>323_-_kaiser_-_clip_1.mp3</flashmp3>

What I can do now is describe to you the atmosphere there in Bell Laboratories at the beginning of the '60s. I was hired into the acoustics and speech research area. My first department head and director was Ed David. Actually, the evening of the first day that I interviewed at Bell Laboratories was the day that Ed was promoted from department head to director. They had taken me out to dinner that evening and at the table across the way at the same restaurant was Ed David and two other people; they were celebrating his promotion to director. Of course, Ed eventually became executive director when John Pierce retired. Ed's office door was always open and he was always available should you have any questions or concerns.

Acoustics research; "cocktail party effect"

Kaiser:

Some of the very first work that I did after I started at the Labs—at the suggestion of Ed David, in fact—was trying to model what's called the cocktail party effect. Now the cocktail party effect is the following: You are in a room where you have a number of different people speaking and talking in different conversations. If you are in a small group and the conversation is not terribly interesting, you might hear a couple of key words over in one of the other conversations that is more interesting to you. Then you focus and try to understand what's being said on the other side of the room. You try to follow that conversation through all of this so-called babble, all the other words and different conversations that are going on and constituting the acoustic field. So the problem was simply how to take the signals from two microphones and process them so that I could get the same improvement in the sound signal out of those mikes that a person is able to achieve when he or she is focusing on a distant conversation at a cocktail party. Now if you use just the normal linear antenna array theory, you can only do about 3 dB improvement. Humans get about 10 to 12 dB improvement, hence the challenge.

Goldstein:

Can you tell me how you measured intelligibility in terms of decibels?

Kaiser:

Well, there are certain tests that you run, checking whether subjects heard words right or not. I don't remember all those fine details. The important thing was could I build some apparatus to reproduce the effect. So then I had to begin to think about how the ears are working and what is the significant information in the speech signal. I was eventually able to build both an analog system that would do nearly as well as the human, and a simple digital version of the same thing, housed in a chassis not much bigger than your recorder there—about a 5-inch by 7-inch chassis—with the electronics available circa 1960. We even got a patent on it, David and I. My first patent at Bell Laboratories was for that system.

Signal processing; digital filter design

Kaiser:

That got me very interested in signal processing. Now, at the time I arrived at Bell Laboratories, a change in the means of doing research in the speech coding area was underway. Instead of the old way, which was to test an idea about a new way to do things by designing the electronics that embodied that idea and then running tests on that physical embodiment, we were starting to simulate the system—a compression system, an encoding system, whatever—on the general purpose digital computer. Then we would just run a test of the new idea, taking speech and running that through the simulated system and listening to the result. It was much faster and more versatile. So I got much more interested in how you took continuous systems and got the discrete models necessary for the simulation. With my control background, I knew continuous systems and filter design quite well, and I tried to carry over some of the same ideas to the discrete world. A lot of it carries over as far as the recursive filters are concerned. These design techniques carry over directly via the different transform techniques, the Z transform, the bilinear Z transform, the matched Z transform, and so forth. But one feature of digital systems is that it's very easy to build finite impulse response digital filters, whereas these are very difficult to build as continuous filters. This was coupled with the fact that right at the time that I arrived there was a marvelous piece of new software that had been developed by John Kelly, Vic Vyssotsky, and Carol Lochbaum called the BLODI Compiler—the Block Diagram Compiler for simulating digital systems. So if you could draw a block diagram for your system, you could usually describe those individual blocks with simple instructions in the Block Diagram Compiler. That would enable you to compile all the interconnections for the blocks that you had and produce a program into which you could just enter your input and get the results out. Now Hank McDonald was very interested at that time in doing the analog-to-digital and digital-to-analog conversions required to couple to the digital simulators. Hank had built—by an external contract, I think—some equipment that we called the HARE gear. I don't remember what that stood for now, but it was basically an analog-to-digital converter. It would write a digital tape at about 40 kilo samples a second, if I remember correctly. Then you could run it back the other way, digital-to-analog. So if you were going to process speech, you'd talk in a microphone and the HARE gear would digitize a tape. You would then take that tape and a deck of punched cards as input to the BLODI Compiler—one card per block in your block diagram—over to the general purpose machine. You'd submit the deck with the tape, and back would come another tape. You went back to the HARE gear and listened to the result. It was really marvelous. So I got very interested in how you take these complicated systems and get rather complex linear filters to realize almost any filter shape that you wanted. My first three years after designing the cocktail party effect apparatus were spent in developing a broad range of filter design algorithms. That was the task that they let me work on, both recursive filter design and non-recursive filter design. The culmination of the writing that I did at that point was a short paper I presented at the first Allerton Conference on Circuits and Systems. Allerton House is a little conference center that the University of Illinois runs about twenty miles west-southwest of Champaign-Urbana. Mac Valkenburg was the person that was most responsible for the Allerton Conference. I always remember that first meeting because I presented my digital filter paper there. I also presented my work on the I0sinh window for that first conference. As a result, a lot of people now associate me with that window, the I0sinh window, which finds use in both digital filter design and spectrum analysis.

Abbate:

What is the spectral window?

Kaiser:

A spectral window is simply a function that you use when you're trying to find the spectrum of the signal in order to minimize end effects. For example, if you are doing a discrete Fourier transform (DFT) of a signal, the section of signal that you are applying the transform to is assumed to be cyclic. In other words, if I give you a segment that is 256 samples long, it is assumed that this signal is really those 256 samples, followed by a repetition of those same 256 samples, and then another repetition, and so on and so on. Now, in reality, you just have a section of signal. It doesn't repeat. It's very important, therefore, to pay attention to what happens where the final sample value is followed by the return of the first sample value when the entire sequence is repeated. So what you would usually do is to take the signal that you've got, e.g. 256 samples, and multiply it sample by sample by your spectral window, which would be a Gaussian curve, or something of that order—that is, a hump shape that becomes rather small at the edges. What you want to do is to try to find the precise shape that you need to get as much energy in the center part, but yet minimize the edge effects. At the time I had gotten to that picture, the window functions that people were using were the von Hann, or raised cosine window, and the Blackman window. Hamming found that if he added a little pedestal to the raised cosine window you could improve the side lobe behavior quite a bit more, so you got the Hamming window, 0.54 + 0.46 cosine π t. We are digressing now into the specifics of windows, but that's okay, because it's important. Right at that time, about 1960-'61, Henry Pollak, who was department head in the math research area at Bell Labs, and two of his staff, Henry Landau, and Dave Slepian, solved the problem of finding that set of functions that had maximum energy in the main lobe consistent with certain roll-offs in the side lobes. Those functions gave a maximum of energy concentration. They showed that the functions that came out of that were the prolate spheroidal wave functions, the solutions of a certain second-order differential equation in physics. That group wrote a beautiful set of papers. They had a postdoc, Jost Sauteur I believe, come in and code up a solution to this differential equation. They also had Estelle Sonnenblik, who was a programmer there at the labs and who worked for Slepian, work on it. The program was about 600 lines of FORTRAN code. I used that program to generate the prolate functions I needed for filter design, and it worked just beautifully. One of the nice features of Bell Laboratories is there were a lot very bright people around, and one of the fellows that I always enjoyed talking to was Ben Logan, a tall Texan from Big Spring, Texas. Ben had an office just down the hall from us. At that point, Ben was doing a Ph.D. thesis at Colombia on high pass functions. So one day I went in Ben's office and his chalkboard was just filled with equations. We had chalkboards in those days, not the magic marker boards of today. Way down in the left-hand corner of Ben's chalkboard was this transform pair, the I0sinh transform pair. I didn't know what I0 was, but when I looked at the transform pair I was struck by the arguments of the I0 function and the sinh function. It looked a lot like sin(x)/x , which is the transform of the rectangular window. But it had an extra parameter, alpha, in there, and I didn't know what I0 was. I said, "Ben, what's I0?" He came back with "Oh, that's the modified Bessel function of the first kind and order zero." I said, "Thanks a lot, Ben, but what is that?" He said, "You know, it's just a basic Bessel function but with purely imaginary argument." So I copied down the transform pair and went back to my office. I looked through a few books to learn more about Bessel functions. The strange part of the transform was the argument of the sinh function. I said "I will just try to compute that function to see what it looks like.” I had been using a prolate window with a parameter of, let's just say, 6.0. Well, I found where the first zero was; it defined the main lobe. I then adjusted the parameter in the I0sinh so that it had the same first zero crossing point. I wrote a program, not more than twelve lines of code—that was all it took to calculate the I0 function using a straight brute-force embodiment of power series expansion. I got the data back and when I compared the I0 function to the prolate, I said, "What's going on here? They look almost identical!" I thought, "Did I made a coding mistake? Did I get the print statement in my program wrong so that it is printing X, X instead of X, Y?" No, it was correct. The answers were within about a tenth of a percent of one another. One program required 600 lines of code and the other ten or twelve lines of code! Then, over the next couple of months, I initiated a short study, a rainy-day game sort of thing, comparing the I0sinh function with the prolates. I found out they were extremely good approximations of the prolate but far simpler to compute. I said, "That's it!"

Digital signal processing; digital filter implementation in integrated circuits

Kaiser:

So that was made a part of my first Allerton Conference paper. The paper also outlined a number of different means for doing finite impulse response filter design and recursive filter designs as well. Then it got into some of the quantization problems encountered including limit cycles. So that first Allerton paper contains most of what someone taking a course in Digital Signal Processing today, more than thirty years later, would need to know. It's amazing, all the main ideas were in there. That paper opened up a lot of people's eyes. This now brings me to a slight digression: I'm one of those people who, when I've done some work, always says "Oh, somebody else is bound to have already done this before." I remember, I asked myself why, with all the work that was done in sampled data systems, with Bill Linvill, Ron Howard and Bob Sittler who were at MIT, and Lotfi Zadeh, and John Ragazzini at Columbia (those were the two big sampled data schools), why those fellows didn't do more with the digital filtering. I think the problem was that at both places, they didn't have the economic means for implementing to any great extent the sample data systems, which were primarily motivated by control system problems— mainly in radar, where you've got periodic signals arising from the antenna scan, the circular scan. After all, the integrated circuit had not yet come along. After the development of the integrated circuit, however, it became possible for people to start thinking about implementation seriously. That's what made DSP take off! Now, once we start talking about implementation, we must return to the role of Hank McDonald. Hank was the sort of fellow who was very much interested in building systems and seeing the new technology become an integral part of the Bell system. Let me tell you how we implemented digital filters. This is the work that is described in the paper published by Hank, Leland Jackson, and myself in the IEEE Transactions on Audio in September, 1968. First, I will explain something about the economics of the integrated circuit. Think of the designer of elements of a communication system: he or she has what I call the big parts box next to his or her desk. The parts box has a couple of different compartments, labeled "cheapest element," "next cheapest element," and "most expensive element." Remember, economics is always the name of the game. The engineer is going to try to use as much of the cheap stuff as possible. Well, for continuous filters design, the cheapest thing was wire. That's the cheapest thing going. Next were resistors and capacitors. Inductors were fairly expensive. The most expensive thing, however, was gain. With the vacuum tube supplying gain, you had to provide the filament supply and the plate supply, called the B+, and so on. Consequently, filters were implemented primarily out of RLCs. Oh, maybe a little later you could get some active filters using miniature tubes, but it was mainly RLCs. These components cost money, so the design techniques for continuous filters were set up so that you were always trying to get the filter of minimum order to meet your specifications. Minimum order meant a minimum number of parts. When we got the integrated circuit, it was like somebody completely changed the roles. The cheapest thing around now was gain—supplied by the transistor. You could lay down the resistors, that was easy. Capacitors were fairly easy. Inductors were still pretty hard, but you didn't really need those. You could leave those out of an instrument. The most expensive thing around, however, was the interconnection—the wire! All of the sudden, the tables were completely turned around. That meant when we thought about design techniques, we didn't necessarily have the same constraints that we had with the continuous filters. Now, Hank McDonald would regularly talk to his friends that were working in the integrated circuit development area. Four times a year, he ran an exercise where he would check in to find out how small they were able to build the circuits, how many elements they were able to get on a chip, where it was going, layout problems, and all that. He could track the economics of using this technology for circuit implementations. So we had a curve, a cost-per-bit kind of thing, which you could actually look at. You knew the cost was coming down steadily. So by 1966, it became possible for us think about building—in hardware—a multiplexed digital filter. There was still a big question, though. We could see how to replace a couple of resistors and a couple of capacitors with a lot of computational circuitry. But was this really progress? So Hank had an idea. There was one other problem with the integrated circuits that I haven't mentioned. The logic elements used in integrated circuits work very fast. They switch literally at rates measured in the megahertz—1 to 2 or 3 megahertz was the speed of the logic we had in those days. And when the engineer puts in any hardware to do an add or a multiply, he wants that hardware to be working all the time, at the fastest rate that it can go. Our application was processing speech, however, which is down at the 10 kilohertz sample rate. So the problem was: how do you make efficient use of this 1 megahertz logic when you are doing operations that are occurring at one-hundredth of that rate? The key idea of Hank's was to multiplex the filter. Hank McDonald came into my office—I remember this very well—one Friday afternoon and he said, "Jim, look, I've got this idea." Then he described to me in detail his idea on multiplexing. He showed me how he wanted to do it, and he was doing serial arithmetic, not parallel arithmetic. That had the advantage of using only one wire with a number of sequential bits on that single line. That way, when you switch things around to do the multiplexing you are only switching one line, which is a whole lot easier than switching twelve lines all at exactly the same time for a 12-bit number. So Hank, after he outlined his multiplexing scheme asked, "What's wrong with it?" I just sat and looked at his scheme and finally said, "Hank, I can't see anything wrong with the idea." We discussed it more, talked over what we thought the properties of the design were, and I said, "I just cannot see any reason why that shouldn't work. In fact, that structure will let us do both recursive filter implementations and non-recursive filter implementation, to whatever order we want. That is an absolutely beautiful structure." So at that point, Hank started to do a detailed hardware design. He had a number of projects also going on, however, so he enlisted the help of Leland Jackson, a student we had with us who was beginning to work on a Ph.D. at Stevens Institute of Technology. Jackson's work at Stevens was on analyzing quantization effects in digital filters, both correlated noise (the limit cycle problem), and then uncorrelated noise—i.e., how you sequence the computations you are going to do to minimize noise when you are building a complicated filter. Under Hank's tutelage, Leland did the hard work of implementing Hank's structure. We completed that implementation in '67 and reported on it at the IEEE convention in New York. That was McDonald, Jackson, and me, but I was by far the least important contributor to the team. I mean, Hank had the basic multiplexing idea, Leland basically built it, and I helped with just a few of the examples and acted as a sounding board. That was my part in it. That implementation paper really got things going. The physical details of that implementation were these: a 19-inch wide rack, about 4.5 inches high and using over 200 dual in-line packages. And today they now get about fifty times that power in one small IC using about a watt of power or less!

Goldstein:

Were those packages off-the-shelf components?

Kaiser:

Yes. Those were the dual in-line packages and maybe you'd get a dozen or so gates on each, or even more than that. Leland discovered the problems of the limit cycles in the hardware and also devised a way around them.

Collaborations with Lincoln Laboratory

Kaiser:

But I am getting ahead of the developments also. I'm going to back up just a little bit here. One of the fellows that I did some very early work with at the laboratories was Roger Golden. Roger was very much into the speech area, and Roger and I had written a program to perform the bilinear Z transformation. We could take any continuous filter and obtain its digital equivalent. Very nice. Now, after we had done this we received a call from some researchers at Lincoln Laboratory. They wanted to come down to visit us to see what we were doing and exchange ideas. The two fellows that came down were Ben Gold and Charlie Rader.

Abbate:

When was this?

Kaiser:

This had to have been around 1963. It might have been as early as '62, but somewhere around there. Anyway, we started explaining to them what we were doing, and I saw their jaws drop. They were amazed because they had been doing some of the same things that we had been doing, but we had gotten a bit further. I think the primary reason for that was that we had the Block Diagram Compiler, BLODI. We had a really neat means for instantly taking our ideas and designs and trying them out on all kinds of signals by simulation using BLODI. They didn't have a good compiler yet to do that.

Abbate:

It strikes me that both the tools you had for computation and the close relationship with the IC people helped.

Kaiser:

Yes. The integrated circuit people were in another wing of the building. Their activity was growing, but again, it was Hank who made the difference. I keep coming back to Hank because he had the vision to bring things together. I realize this even more as I've grown older. He kept his finger on the pulse of developments in the integrated circuit area. He made the effort to go over and find out what those integrated circuit fellows were doing, "How are you fellows progressing? What are your problems? When are you going to have 32 bits on a chip (not 32 kilobits or 32 megabits, mind you, just 32 bits on a chip)? How much is this going to cost?"

Goldstein:

What was his position?

Kaiser:

Hank was a department head.

Goldstein:

What was the formal name of the department?

Kaiser:

In Bell Labs the names of the department changed to follow the work that was going on in the department. Personally, I was probably in maybe eight or nine different departments.

Abbate:

Okay, to get back to your story about Gold and Rader.

Kaiser:

Yes. Ben and Charlie came down, and it was very fruitful for them. They also made some good remarks to us, so that was the beginning of our interaction with them.

Abbate:

What was their application?

Kaiser:

They were doing speech work. There was also radar and other things going on up there at Lincoln Laboratory. But as far as I knew, most of their work at that point was in the speech area. Ben was a great devotee of Homer Dudley of Bell Laboratories and his vocoder work. He was the real inventor of the vocoder. Homer was the grandfather, and then Ben was the father. After Homer died, in the late '70s or maybe in the early '80s, Ben gave a tribute to Homer Dudley at one of the ICASSP meetings. Ben had gotten a xerox copy of two pages out of Homer Dudley's Bell Labs notebook describing the invention of the vocoder. He showed those two VU graphs during his presentation on Homer, and when I read those two pages, I realized I have never seen such succinct writing in my life. On those two pages Homer describes the vocoder system and probably about a dozen different variations—and in sufficient detail that you could go out and build the thing. I mean, Homer just covered the map in only two pages!

Goldstein:

Well, the vocoder raises the same question that Janet was getting at a second ago. It's easy to understand the purpose of Bell Labs' vocoder research. What was the reason the they were doing that work up at Lincoln Labs?

Kaiser:

That's a good question for Ben and Charlie. I remember military applications, such as secure communications, low bandwidth communications, and the like. All those require a very good understanding of what constitutes information and speech. Charlie and Ben wrote the very first textbook on DSP, from about '66 to '67. It was a landmark event.

Publications and committees

Kaiser:

Let me review my own writing: there was my first Allerton conference paper, and two years later, in '65, the third Allerton conference paper where I spoke mainly about quantization effects. Then in 1966, Frank Kuo and I edited a book called System Analysis by Digital Computer, and we had a chapter in there on the Block Diagram Compiler and its floating point additions. I wrote chapter 7, the one on digital filters, which was essentially my first and third Allerton papers together, with a little bit of other information. That chapter got xeroxed and reproduced a number of places. It was right at this time though that I was also starting to go around and give a lot of talks. Electrical engineering departments were inviting me to speak because they were always interested in getting people from Bell Labs. They wanted to hear about the latest things that were going on at Bell Labs, what kind of problems our graduate students should be working on, promising thesis areas, and so forth. So we at Bell labs had some influence on where the developments were going. That was a wonderful thing. Also, I was invited to serve on the COSINE committee. This was a committee that was set up by the National Academy of Engineering under NSF support. We worked out of D.C. It was mainly electrical engineering department educators: Lotfi Zadeh, Sam Seeley, Mac Van Valkenberg, Bill Huggins, Bill Surber, Taylor Booth, and Ben Coates. I'm probably forgetting a few names, but there were ten or a dozen of us. The real problem we tried to tackle was how electrical engineering departments should come to grips with the role that the digital computer was playing. We looked at questions such as whether the EE departments should manage their central computing facilities themselves; how they should structure courses to teach students how to build systems that use digital technology; and how to use computers as a design and engineering tool in courses, in the same way that students learn how to use MATLAB and Mathematica now. There were these three major areas. There were two different things we did. About every eighteen months, we'd get electrical engineering department heads together in a meeting. There were perhaps 200 or so electrical engineering departments across the country at that time. We typically would get the heads from about 120 or 130 of them to come to a two-day meeting with talks and so forth on these three different problems. The other thing we would do was to get together maybe eight or ten people and have a workshop on what should constitute a basic curriculum for computer scientists, what constitutes digital laboratories, and so forth. We'd work for a week, bang these ideas around and come up with summary reports. I served on that committee from 1966 until about '72 or '73, serving mostly as the secretary. I helped them organize meetings. It was a wonderful experience for me because even though I was still a very active researcher at Bell Laboratories, I got to meet and work with these other academics. I can remember right now, in fact, that at one of these meetings asking Lotfi Zadeh—who was out at Berkeley at that point—"What are you working on now?" It was fuzzy sets! He described it to me, and he said, "Well, here's my first report." I still had that report, and I gave it to Paul Wang at Duke because Paul was very interested in fuzzy sets. Paul was quite happy to get that "first edition" of Lotfi's work.

Design methodology; education

Goldstein:

Let me come back to a different question concerning the Block Diagram Compiler. Was this developed as part of a larger strategy of simulation at Bell Labs? Or was that work and your work on digital filters conceived more or less independently?

Kaiser:

The work was done independently. If I had a strategy, it was trying to carry over as much of the available knowledge as possible from the continuous filter area to the discrete area. But then that was mainly all the knowledge about recursive filters. In the late '60s, after we had done our first hardware implementation, Hank tried to envision how this was going to affect how we build the communication systems. We sat down and we tried to figure out how long it was going to take before these ideas really started to permeate the physical plant. We made some prognostications about how long it was going to take. In fact, I gave a talk down at the University of Florida in Gainesville in April 1971, and I ended the talk on hardware implementation with a prediction of where we expected things to go. They made a video tape of that talk. About ten years later, in 1981, I returned to Gainesville with Don Childers as my host to give another talk. Don said to me, "Incidentally Jim, we've shown that videotape of your talk ten years ago in every class that's taken DSP at the University of Florida." I said "Gee, that talk is way out of date." He said, "Sure, but, when I play the tape for the students, I don't tell them when it was made. Everything that you said was going to happen, did happen. It's as if you were clairvoyant. The only thing, it didn't happen as fast as you thought it was going to happen." The real problem was education. A few of us at Bell Laboratories and a few at MIT and a few other places across the country were coming to understand the new methodology. But before it could get into the physical plant, that methodology had to be understood and be commonplace amongst the broad range of designers. The professors in the universities had to understand it first and then instill those ideas in their students; we really hadn't factored that educational time lag into our prognostications. When we did, that added an extra four or five years for DSP to begin to take off. Now, you asked about whether there some grand pattern in the design methodology. The thing was, there were old timers—especially in industry—who had been designing filters, continuous filters, for years, and all of the sudden these fellows were told, "Now look, I want you to build digital filters." Their reaction was, "I don't know what a digital filter is." They were completely lost and some of these fellows didn't want to learn the new stuff. They said, "There's too much new to learn! Don't bother me with the new stuff." So one of my goals was to preserve all the knowledge that those fellows already had and say to them, "Look, all you've got to do is run this little program with that knowledge you already have and it will design a digital filter for you. You know how to do it!" That's basically what the bilinear transformation does. It carries the continuous filter designs over to the discrete design. You see, if you formulate the specifications for a filter in the discrete domain without looking at the continuous domain, the specs will be the same as they would be for the continuous domain filter. If you could design a continuous filter with a bilinear transformation, that design that you got by trying to solve the problem in the discrete domain would just be the design in the continuous domain transformed. In fact, some fellows who were doing things totally in the discrete domain would say, "Oh, we can design this kind of classic filter and that kind of classic filter," and write a new paper on it. They didn't get it. Their work was simply the original continuous filter with this transformation. That's all. This transformation technique made some of these old-timers feel a lot more comfortable. It also enabled us to bring a tremendous body of filter design know-how into the discrete domain.

Transition from analog to digital; multiplexing

Goldstein:

So did these techniques bring an end to developments in conventional analog filter design? Did the field stagnate?

Kaiser:

What was happening in the continuous filter design business was that people were becoming much more interested in active filters. Remember, with the transistor and the integrated circuit you have that cheap gain available. We didn't need the filament supply, and so on, to get gain. So there was interest in the active RC filters. For example, we followed very closely the work of Ken Laker, who worked with active RC and then switched capacitor technologies at Bell Labs. Digital filters interact with the continuous world, so when you are working with them, you've got to get involved with analog-to-digital conversion, and you've got to worry about the aliasing problems. You're going to have to do some filtering in the continuous domain to quasi-band-limit your signal before you sample. Most of the implementation of those band-limiting filters was done with active RC. So I worked in that area also. In fact, one of my functions, in addition to my research, was to look over designs for the fellows in our sister departments who were building hardware. They would come to me, and it was good because it would help test my design programs out, and it let us know what problems those fellows were having that we didn't have software to solve.

Abbate:

What were the first pieces of hardware that you actually used? At some point you are actually implementing digital filters in hardware for use in the telephone system, right?

Kaiser:

<flashmp3>323_-_kaiser_-_clip_2.mp3</flashmp3>

Well, there is an economic issue related to the applications in the Bell system. In the telephone system, you have copper wire connecting you to the central office and one central office to another. You'd like to use the copper wire as efficiently as possible, so that means you'd like to get more than one phone conversation over a single pair of wires. So multiplexing was crucial. The early multiplexing in the physical plant was all analog; it was N-carrier. You would modulate a voice channel up to the 60 to 108 kilohertz range—that's twelve 4 kilohertz channels. So you used many modulators and band filters. Now, they would use those carrier systems in the interoffice trunks, the wires that connected one central office to another. When the digital circuitry started becoming widely available, we said, "Let's take those signals and instead of frequency-division multiplexing them, let's time-division multiplex them. What we'll do is convert the signal to digital form, pulse code modulation (PCM)." With a 4 kilohertz band for speech, you would sample the speech signal at 8 kilohertz, using 8 bits per sample; that's 64 kilobits per second for each speech signal. One could send 1.5 megabits per second along this copper wire, however. Hence, using time-division multiplexing, we could send twenty-four speech channels simultaneously on that single wire. Now, you have to put repeaters about every mile, because at 1.5 megabits per second these individual pulses might start out as 0.67 microseconds wide, but after a mile the dispersion in the cable broadens the individual pulses to about 20 microseconds. The preamplifier in the repeater would concentrate that energy, decide whether there was a pulse there or not, and if there was, it would generate a new, clean 0.67 microsecond pulse. By the way, one of the fellows who designed those repeaters was John Mayo, a member of the technical staff, later a department head, and eventually president of the Bell Laboratories. So digital hardware in the interoffice trunk came a little bit ahead of the digital filter implementations. But then once you go digital somewhere in the system, you say, "Now I can do all these other operations digitally, all these complex pieces of equipment that sit in the central office, such as the touch-tone detector systems." So for our first hardware implementation, we showed that we could use this multiplexed filter hardware for realizing the touch-tone receiver. The touch-tone receiver is that little box in the system that decides from the pair of tones that get generated when you've pressed any of the buttons on your touch-tone set. Do you know how that works? When you press a button on your set, it generates two tones simultaneously—one corresponding to the row frequency and the other to the column frequency of the button location. The equipment on the other end has to listen to these pair of tones that are coming and decide right off which tones they are and, hence, which button was pushed. I think we needed about fifty second-order digital filter sections to get the touch-tone receiver to do all that. There were four row frequencies and four column frequencies. The band-pass filters tuned to each of these eight frequencies were second-order. There were also two band separation filters and a bank of low-pass smoothing filters. By the time you put it all together, it was comprised of maybe fifty filter sections. We did that all in one multiplexed second-order filter section. That was a very good example of the use of Hank's multiplexing scheme. It was a beautiful example to be the first. There is more to this. This is a story that Hank would be able to tell you in much more detail, but when you've got a development such as this, you start to get into problems concerning manpower and the education process. The department head, or the director, who is going to be responsible for the next generation of multiplexing equipment, has got to work with his designers. Now, he's got a schedule to meet and a budget to meet, and if his designers don't know the new technology, then he's got to go with the technology that's going to enable him to meet his budget on schedule. That was the situation, even within the Bell system, that delayed our getting the digital ideas into the system even sooner. And, of course, you always have personality and turf problems that also come in. But we were shielded from those problems in the research area. I think the really good research managers are the ones who get key people, just let those key people work, and keep the wolves away from the door. Someone who lets the researcher think and work as broadly as possible. Remember I said I went to Bell Labs for real world experience? Well, it wasn't quite the real world. It was a marvelous world, and marvelous things were done there. That was because we were shielded from much of the petty stuff that comes up. We could really work; we interacted a lot with our colleagues. When management became tighter and there were heavy constraints in the budgets, there was strong competition among departments which then tended to cut off communication between researchers and their colleagues. That was not as good. You want a much more open environment for effective research. That was Bell Labs, and that was what was so marvelous about it.

Radar applications

Goldstein:

So far, you've talked mostly about signal processing in communications applications. I wonder whether your work was of equal interest to people working in radar or perhaps other areas? Did you have any direct communication with these people?

Kaiser:

Oh, yes. I never did much with the radar fellows, but another area that just turned out to be huge was geophysics. At one point, there was more signal processing going on in Tulsa, Oklahoma than there was anyplace else in the world. They were interested in geophysical prospecting for petroleum reserves.

Abbate:

People such as Sven Treitel?

Kaiser:

Right. Sven was at MIT and Enders Robinson was also at MIT in the early '50s. I was there at the same time they were there, but I didn't really know them at MIT. That didn't come until later.

Abbate:

Was there a connection between Bell Labs and them?

Kaiser:

Bell Labs? Not really. When we started publishing a lot of the things that we were doing, the fellows down in Tulsa showed keen interest. So people such as John Shanks and Sven Treitel started coming to some of our workshops like the Arden House conferences.

Emergence of digital signal processing field; Arden House workshops

Abbate:

Can you tell me about how DSP emerged as a field? If there were people in these different areas interested in the work, then how did that come together?

Kaiser:

Here's how. We had a loose connection with the fellows at IBM and the fellows at MIT. Then, when the fast Fourier transform (FFT) came along with Jim Cooley and John Tukey. Jim was at IBM, and Tukey was an associate director at Bell Labs, a statistician who was John Pierce's right hand man. There was a workshop at Arden House that was organized by some of the fellows in the IEEE audio and Electroacoustics group. I was only up there one day. Following the workshop on the FFT in 1967, that group organized a technical committee on digital signal processing, which they asked me to join. That technical committee was Al Oppenheim, Larry Rabiner, and Ron Schafer, and a couple of other fellows from Bell Labs including Ron Crochiere, Leland Jackson, Howard Helms, and Dave Bergland. Al was up at MIT and Cliff Weinstein, Ben Gold, and Charlie Rader were at Lincoln Laboratory. Ken Steiglitz was there from Princeton, and then Jim Cooley and Harvey Silverman were there from IBM Yorktown Heights. There was also Russ Mersereau, Joe Fisher and Steve Lerman. The committee was a very significant force in DSP. It got together every two months; two times in the United Engineering Center in New York, at the IEEE's offices up there on the tenth floor, and the third time it would be up in the Boston area, and then this cycle would be repeated. It was convenient because you could take the shuttle up to Boston for the day and return in the evening. It was at those meetings that we planned the IEEE Press books, the program book, and the selected papers on DSP books, and we planned the Arden House workshops—we organized and ran three workshops there that came about two years apart. We also did the normal technical committee work, which was reviewing the papers for the annual ICASSP. Moreover, when we'd get together we would interact with one another. You know, when you're out at lunch or chatting on the side while something else was going on, you would exchange information about what kinds of things you are working on. There was a lot of interaction then between the IBM fellows, the fellows at MIT, us at Bell Labs, and Ken Steiglitz and others down at Princeton. There was a lot of close work there. Those Arden House meetings were invaluable, because they brought people from all over the country, and even overseas. Schuessler from Germany came to almost every Arden House workshop, and there were also fellows from Switzerland: Federico Bonzanigo and Fausto Pelandini, and those working in Professor Bauman's technical physics group over there at the ETH. They would often bring their best students along, too.

Goldstein:

Does the importance of attendance at Arden House mean that there was a knowledge to be communicated there that couldn't be transmitted effectively via papers?

Kaiser:

There was the cutting-edge research, and the fellows who came there were trying to say, "Look, this is where we are. These are the kinds of things we're doing, and this is what we've found." There was tremendous interest in speech processing, so the speech areas were very large. Within speech, there was the hardware implementation, design techniques, quantization effects, and so forth. Those were the central topics.

Music applications

Kaiser:

One of the application areas that evolved out of this was the application to music. You could use this beast to generate new sounds or to reproduce organ, piano, and all the different instruments. This has been a huge factor for the engineering departments in getting people interested in DSP. Max Matthews was significant there. Did Larry Rabiner talk much about Max in your interview with him?

Goldstein:

No, but I've heard a lot about Max from John Pierce.

Kaiser:

John Pierce was our executive director. That's three levels up. One of the beautiful things about Bell Labs' organization was that there were very few levels of management and they were all technical people. There was the researcher. Above him was the department head who looked after six or seven researchers. The director had three to five departments. Then there was the executive director, then the vice president in charge of research, and that was it. Above him there was only the president of Bell Labs. When Pierce was the executive director, he knew all of his researchers. Baker, who was vice president in charge of research, had 500 Ph.D.s and he knew every one of them by first name! When I saw him walking down the hall, he would say, "Hi, Jim." Not only did he address you by name, he usually knew exactly what you were doing. Think how unusual it is for a polymer chemist to know what's going on in digital signal processing for speech and communications. You name it, Baker knew it. He was a wonderful person. One other thing: he spoke in complete sentences. He was one of the few people who could do that. You would listen to him talk and you would hear him start out a sentence and you'd think, "Oh, he can't make that into a sentence." Then you would listen further and, boy, he would figure it out somehow and it would come out a complete sentence; a long sentence, perhaps, a complex sentence, but a complete sentence. Max Matthews was in the acoustic research area, but Max did the music compiler. He was an amateur violinist, so Max was very motivated by music. That was the thing that kept him sane, I think. He compiled instruments using something akin to BLODI. It compiled the instruments, put the instruments together and the notes together. After he retired from Bell Laboratories he went out to CCRMA at Stanford, where he is still very active.

Interactions of West and East coast researchers; aerospace applications

Abbate:

Speaking of the West Coast, what were the aerospace people doing in signal processing?

Kaiser:

Those fellows were building digital control systems. Stan White was one of the main people out there. Stan was at North American Rockwell, and Stan's background was in control. Of course, a lot of it was classified because it was military. It was either classified as secret or it was company confidential. So you could never really gather what on earth they were doing.

Goldstein:

Did that keep them out of the workshops?

Kaiser:

They would still participate in the workshops, but you couldn't really find out where their cutting edge was. There appeared to be a broader gap between what they told you and what was really going on.

Abbate:

They weren't putting in as much as they were taking out?

Kaiser:

There was a time lag. They were putting in, but what they would say was, "Yes, we were working on that a couple years ago. Is that where you fellows are? Back there?" Remember, though, there weren't too many places on the West Coast where there was cutting-edge work done in engineering departments. One of the problems that we had was that our DSP committee was mainly composed of East Coast people. The East Coast group could get together easier. You could fly to Boston in one hour. The West Coast was five or six hours away. That's a lot of time. So if we had somebody from there on the committee, we didn't see him every two months. We saw him twice a year if we were lucky. The geographic separation was, in the early days, a bit of a handicap. But then Stanford became more important. Bernie Widrow went there, for example. And there was that exodus from Bell Labs in the late 1950's, before I got there, out to Berkeley: Charlie Desoer, Don Pederson, Ernie Kuh, and later David Hodges and Dave Messerschmidt. Those are the names off the top of my head. I knew Messerschmidt at Bell Labs. The three big people, though, were Desoer, Kuh, and Pederson. Those fellows had been in exploratory development at Bell Labs. That was a research group within the development area that Bell Labs had set up to do development that was right at the very forefront. It was distinct from the research area, though. Then one day one of the top management at Bell Labs suddenly said, "What are you fellows in the development area doing research for?" So when these fellows could no longer have the flexibility that they wanted, they made a migration to Berkeley. Those three were marvelous minds, first-rate circuit guys.

Defining signal processing and DSP fields

Goldstein:

I have a question about the term DSP. Once people started talking about "digital signal processing," did that retroactively define "signal processing?" Was signal processing a coherent field before the mid-'60s?

Kaiser:

We never really thought about it as signal processing. In electronics it was maybe signal generation or signal shaping or that kind of thing. I can never remember us saying, "We'll do it by signal processing." But then once you've gone digital, now also you realize, "Gee, stock market prices are digital signals. They are sampled signals. A sampled signal from a musical instrument, a music signal, has something in common with them" It's not just this little communications thing that you were working on, or this radar thing. It's very broad. The operations you are doing are, essentially, the same kind of things in each case. And the development of computer processor power is what prompted that realization. It was the tool that allowed you to do things with software on signals no matter where the signals came from, not just a particular problem, such as an electronic circuit that's doing a sweep on an oscilloscope or a speech signal.

Goldstein:

Were there any problems consolidating people from different fields into this new field called signal processing? Did a numerical analyst working on stock market data initially feel that he or she had anything in common with a communication system designer?

Kaiser:

I have a personal story about that, concerning my interaction with Hamming. Dick Hamming wrote this book Digital Filters, right? He is best known for his book Numerical Methods for Scientists and Engineers. Where did that book on digital filters come from? Well, Hamming was a numerical analyst who solved real problems. He was very focused on doing things—that is, making a difference by solving real problems. He was trained as an applied mathematician. Before he came to Bell Labs he was actually working at Los Alamos, running the computing machines at the time of the first Alamogordo test. In the '70s, he was hearing all the things we were saying about "digital filters." So he wanted to find out if what we were doing was any different than what he was doing when he was working with difference equations, integrating differential equations on the digital computer. So he started having lunch with me and asking all kinds of very direct questions. When Hamming wants to get answers, it's not for you to ask questions. So we were having lunch together maybe three days out of the week for a half dozen years. I got to know Dick very well. The outcome of that was he tried to get me to write a book on DSP. We were talking about all these things and, in essence, he wrote the book. His digital filters book is largely derived from our lunch conversations, but with his slant on things. I think that Dover is going to release the third edition of that book in their classic reprint series. He's written some eight or nine books, but that mind of his is just totally focused on problem solving. I gave a talk at the tail end of the '60s, or maybe around '70, on digital filters to the mathematics section of the New York Academy of Sciences. I asked myself, "How do you talk to these mathematicians about digital filters?" So what I did was to take the specifications for filters that we were building, design the methodology, and normalize everything to be from zero to one on the abscissa and zero to one on the ordinate. This way, I could give them an idea what the specifications were like for our filters. And these fellows understood that. Here again, it was the difference between theory and practice. The important question for the mathematician is: is there a way to do something? Is it do-able or is it not do-able? Does the solution exist or doesn't it exist? Those are the big questions for them. But for us, it's "Well, we know a solution exists. We want to find a solution and we want that solution to be best in some sense. And we want to be able to build it now." I mean, don't tell me that we know that the Fourier series converges. I want to know how many terms I need. That's the critical question. The window approach gave me a beautiful answer to that question as to how many terms I had to use. It was exactly the tool that I needed to answer that question and find the actual filter coefficients which are also simply the impulse response of the filter itself.

Goldstein:

Did any important contributions come from the mathematicians?

Kaiser:

This might be an oversimplification, but I think that the most important parts of the mathematics that we used were all done prior to the 1930s. Of course, Gauss and Euler and Lagrange and others did the real work that laid the groundwork for what we do. The mathematics that we use is basically theirs. Remember that paper by Heideman and Burrus on the history of the discrete Fourier transform? They found the DFT in Gauss's notebook. It was there. So were those other mathematicians that have done it. Now, the numerical analysts have done a tremendous amount of really good work in computational fluid dynamics, which is a lot of digital signal processing in two or three dimensions. That's highly mathematical work, such as building the electronic wind tunnels—that is, modeling a real wind tunnel. You can do it all by computation.

Abbate:

When was that? Is that a later development?

Kaiser:

Yes, it didn't occur in the '60s. That occurred only after there was a bit of horsepower in general purpose computers.

Abbate:

And by that point, did there exist established signal processing techniques?

Kaiser:

Right. That work uses a few of the filtering ideas that came from DSP. But it's a little bit different, though, because most all of that work is solving partial differential equations on computing machines by numerical methods. It's related, but it's the application area which drives the work and methodology. That was fluid dynamics, something like a cross between speech and geophysics. Actually, there's probably more connection between computational fluid dynamics and geophysical prospecting.

Goldstein:

I'm glad you brought that up, because I wanted to ask whether the development of off-line applications is a different story than those that have to be done in real time.

Kaiser:

Yes, undoubtedly. Another thing that I haven't touched on at all is error correction. I can't say much about that. Hamming is a great person for that. I just think it's marvelous, though, the tremendous level of technology that enables us to, say, receive transmissions from a space probe a billion miles away, transmitting pictures back from planetary fly-bys at only 5 watts! The error correction in the coding for that transmission is tremendous.

Goldstein:

So you were never too involved with that stuff?

Kaiser:

That's right. Not at all.

Optimization

Goldstein:

All right. It sounds like we're suggesting a picture where a core set of techniques find use in a variety of applications which then, in turn, raise new problems of their own. For instance, digital-to-analog conversion was a crucial piece of the puzzle for the kind of work that you were doing. Can you identify any such new problems that came up as a result of the application of some of the core techniques?

Kaiser:

There was a tremendous emphasis on optimization, finding the lowest-order system that would satisfy the set of specifications. More complicated systems tended to be higher order, with a lot more variables you were wanting to adjust or change. There were also related problems, such as the traveling salesman problem, where you try to find the route to send your repairman or your delivery man out on a delivery if you want him to use the minimum amount of time and gas. Then you had Karmarkar algorithm for solving some of those problems. These were important to the Bell system for routing telephone conversations over a network. That work was done in the math research area at Bell Labs; the area where Ben Logan was. You know, when I talked to Ben about the I0sinh stuff, I wasn't asking him "Ben, do you have a window?" I was just looking for a transform pair because I was using these functions for doing my filter design. But Ben worked a lot on a solving these transmission problems. So you see, the opportunity for interaction at Bell Labs helped a great deal. Ken Steiglitz once asked me, "Where did the term "digital filter" come from?" He thought that maybe we had coined it in the early '60s. It turns out, however, that it appears in some papers as early as 1952, I think. Herb Salzer, who had looked at discrete ways to do differentiation and integration for his thesis at MIT, published an IRE paper with the term "digital filter" in it. I think that might be the first place where I encountered that terminology.

Spectrum windows

Goldstein:

Should we talk more about your work on spectrum windows?

Kaiser:

Well, there's a little more of a story there, but almost all that work was done prior to 1963, when I first published it. There's a little moral to this story, and it has to do with where and how you publish ideas that are useful. I published the I0sinh as a first Allerton conference paper, republished it as chapter 7 in the book System Analysis by Digital Computer I edited with Frank Kuo, and then brought it up again in my third Allerton Conference paper. I thought, "There it is. I let the world know about it, I talked about it." But with talks you say the word and then the word is gone. It disappears. It has no lasting archival effect. I reviewed a paper internally at Bell Labs in the early '70s, in which the author was talking about designing FIR filters by the window function method. He commented with regard to windowing that yes, there was this I0sinh window, but it was too difficult to compute. I read that and said, "Where is this person coming from? Don't they understand the simplicity?" So I went around and talked to the person. I said "It takes less than ten lines of code to do this. What do you mean complicated?" "Oh!" he said. "That's all there is to it?" He called me up the next day and said he had programmed the algorithm on his HP-67. It was at that point, in 1973 or so, that I wrote another conference paper. This was for the IEEE Symposium on Circuits and Systems, ISCAS '74. It was a four-page conference paper, and it talks about I0sinh and lays out FIR design by the window function method. It's very complete, all in just one little four-page paper. That made a little more of an impact, but still, I think when engineers read "Bessel functions," they don't want to know about it. They don't want to know about Bessel functions at all, unless they're in electromagnetic theory. Then it's second nature to them, but otherwise, their attitude about Bessel functions is, "I'll take your word for it, but don't bother me with the details." So people don't use it, simply because they haven't taken that extra little five percent effort that's required to understand what's going on. I call this the Bessel function syndrome. I have given a number of talks on window functions; I just show them how simple the Bessel functions are, and compare them to the sine and cosine. They're almost as simple as the sine and cosine functions. There's very little difference in the power series. With cosine, you know, the amplitude just stays fixed and oscillates forever. I0 behaves like a damped cosine.

Goldstein:

Does this story have a happy ending?

Kaiser:

Well, yes. It's in most textbooks now. But still, the authors don't realize the second feature of that window is that it lets you derive the relationship between filter order and filter specifications. It lets you derive that relationship very simply, and in such a way that it is very easy to keep it in your head, so if somebody gives you a set of specifications, in less than a minute's work, you can work out in your head exactly how complicated the filter is going to be. You can get it within ten percent easily, and most of the time within less than five percent. That's the kind of information a designer should carry around in his or her head, i.e. something that is simple but powerful and gets you right to the core of the main first-order relationship.

Speech processing and sound production modeling; nonlinear signal processing

Abbate:

I'd like to hear about some of your later work.

Kaiser:

Fine. I literally spent about twenty years doing both continuous and discrete linear signal processing at Bell Labs. I always had an interest in speech processing and the modeling of sound production in the vocal tract. Being a singer, I was very much aware of sound production. The thing that always drove me up the wall was that my experience with sound production as a singer didn't jibe with the mathematical models that my colleagues were producing at Murray Hill. Now, at Bell Labs, we had gone through a reorganization in the middle of the 1960's, and our Speech and Acoustic Laboratory became two different centers. I went with the center that worked more with the digital computing machine. That's why my career essentially went DSP and not speech. But now, I was coming back full circle and becoming very interested in speech production again. I had a colleague whom I knew from graduate student days at MIT named Herb Teager. I got to know Herb in 1954. He was also in automatic control, and his thesis professor was Don Campbell. Herb got married in June of 1953, and I got married in June of 1954 to his wife's sister. So I had gotten to know Herb fairly well while I was still a student. He finished his degree in '56 and then joined the Navy. He came back to MIT as an assistant professor in '59, just at the time I went to Bell Labs. He was at MIT for awhile, and then he moved over to Boston University. Teager was brilliant. He had a first-class mind and was an unusual person. He was a very good theoretician, and a very good experimentalist. He was very good with his hands and with tools. He had a very good analytical mind. I would see Herb maybe once every several years, and I'd ask him what he was doing. It turned out he was working on trying to understand speech production in the vocal tract. He said, "There's no way that the acoustic signal was generated by an essentially linear model." There was a lot more going on inside the vocal tract that contributes to the production of the signal outside than was included in the models of my colleagues at Bell Labs at Murray Hill.

Goldstein:

Had they deluded themselves or were they using the linear model for the sake of simplicity?

Kaiser:

Well, I think both.

<flashmp3>323_-_kaiser_-_clip_3.mp3</flashmp3>

There are two totally different agendas that underlay Teager's work and that of my colleagues at Bell Labs. All the work that has been published in the speech area was underlain by what I call Agenda One, which is the engineering agenda. The goal there is this: we have got the speech signal outside the mouth, and we can measure it with an electret microphone, or a carbon button microphone, and then transmit that signal down a pair of wires. We can measure that signal to arbitrary accuracy. What we are after is some kind of mathematical model that we can manipulate, that will generate the signal that we measured with the microphone to arbitrary accuracy. It's totally irrelevant whether or not that model bears any resemblance to the physics of production. It only has to be a computationally efficient and adjustable model. That's it: computationally efficient and economically viable, so as to allow one to build the hardware to generate the speech signal as part of the system. That's Agenda One, and it has dominated most all of the work on speech production modeling. Let me put it this way. The signal that is picked up by the ear is basically an oscillating pressure wave. The ear essentially becomes a spectrum analyzer, so you are going to try to get a quasi-linear model for this. If you have a linear model for speech production, then it makes sense for the ear to be essentially modeled as a spectrum analyzer. Then all of this speech production work, the literally thousands of papers written on the various aspects of this, can all be concentrated on this as a linear model. Because almost all this work on modeling was done by electrical engineers, they like to look at things as filters, as block diagrams that have "input," "system," and "output." The "source," or input is the vocal fold oscillation. The filter is represented by the cross-section area of the acoustic tube, and the "output" is the pressure wave at the mouth. That's the filter model and its many variations. That's the approach that was used. On the other hand, Teager looked at that model and he said, "No, it's not right!" When he started going inside the vocal tract and measuring what the air was actually doing when we speak or sing, he found a totally different set of phenomena than what the linear model says should be going on in there. So he writes it up as a paper, writes up proposals, does the whole bit, and the work gets shot down. His proposals are reviewed by the fellows who are working on the engineering agenda. He's working on a different agenda: Agenda Two—the scientific agenda. He says, "Look, let me get the physics right first. Then once I understand what's physically going on in this generation, then I will worry about the mathematical modeling after that, because then I will have much better guidelines as to how to do the modeling and which approximations are meaningful and which ones are not meaningful." So Teager works on Agenda Two. He writes his papers on Agenda Two, but they get reviewed by Agenda One fellows, and the papers are shot down. They all go back to the same people, which, in some cases, were my colleagues at Bell Labs. I only became aware of Teager's difficulty in 1980, when I had gone down to a family get-together in Washington, D.C. Herb and his wife were down there and I was down there with my wife and children, and I started talking to Herb. He was saying a few things to me, and all of a sudden I thought, "Ahah! That's what's going on." Certain things became much more clear to me about certain problems that I had (problems that had come up through my singing, which I was doing very actively). So after I drove back to New Jersey that night, I was up for about another two or three hours writing things down in my notebook. I switched from being a linear person to a nonlinear one literally that night, and that was the end of my linear DSP days. I became a real devotee of Teager's approach, but Teager wouldn't tell me the details of a lot of his work, only the broad outline. I think the reason why was that it was my colleagues who were giving him a hard time. He wasn't about to give me the goodies, so to speak, only to have me spread it around for free at Bell Labs. No way! Absolutely no way! He told me part of it, but as for the rest, he just didn't know what my response would be in this whole business. So it took several years before I could prove my credibility enough so that he would begin to open up. Another thing is that at about that time, I started working in a different area, because of a reorganization at Bell Labs. I was working on something a little different than acoustics. At Bell Labs in 1983, I wrote a technical memorandum on this other way to look at the speech production problem. I got some very interesting reviews! There were conflicts throughout the laboratory. That's another whole story in itself, but it made life a little uncomfortable for me there. So anyway, at the time of divestiture I chose to go to Bellcore, where I could continue with this work.

Abbate:

And where did that lead?

Kaiser:

It turned out I was going to have to do a lot more experimentation of the sort that Teager was doing, but that was not really my strongest point. My strongest point was the signal processing. So I started to concentrate on the nonlinear end of signal processing. I also felt I should begin to write, to let people on the outside know what the problems were. So I started giving talks on the fluid dynamics of the vocal tract, and by now I've given more than fifty talks on this subject. I also started developing nonlinear algorithms for extracting the modulations present. Herb wouldn't tell me the details of his algorithms, but he gave me enough broad ideas so I could put together something that was nearly equivalent to what he was doing, or at least part of what he was doing. A lot of things he did, to this day I don't know how they were done. A very sad event occurred in 1988. I'd usually call Herb up on Friday afternoons and talk with him on the telephone. I called him this one Friday and he said, with a very serious tone of voice, "Oh, just a minute, Jim. I can't talk with you right now. Something's come up. I'll get back to you." He called me back late the next week and said, "I've got some sad news." He said, "The doc had just come in and told me that there was spot on my lung. They did a biopsy, and I just found out I've got lung cancer." I didn't ask him what the prognosis was at that point, but in another two weeks, he was beginning to come to grips with it. He told me, "They gave me twelve to fifteen months." That's really tough. At about that time, he had been invited to give a talk at a NATO Advanced Study Institute planned for August 1989 in Bonas, France, which he accepted. He was beginning to work on preparing a talk as he was going through chemotherapy and radiation. Herb and I were on very good terms at that point. I had invited him down to Bellcore in 1984, and he'd given us a fine seminar on his work and ideas, which I transcribed in detail. It was a very good talk. I had invited a number of fluid dynamicists also to hear the talk, and that was very good. So Herb took that transcript of his talk, added another twenty-five percent new material, and used that as his 1989 Bonas paper. His health was declining markedly, however, so he was unable to go to Bonas. He asked me to deliver that paper for him—it was one of the hardest things I ever did. That talk was received fairly well. But again, it was like David going into the lion's den. You know, many of those fellows didn't want to hear it. Some of them were open-minded about it, but it was like putting your arm down in a bucket of water and pulling it out enough to see what the effect is afterwards...and there's not much. There might be a few ripples. But the important thing was, we jarred people enough to make them begin to think about the problem. The signal processing work turned out to be the big key for me, because I first started writing about the energy separation algorithm. Let me explain something. The signal that is recorded is the pressure wave, so all of the modeling concentrates on generating that pressure wave. But the thing is, this is a wind-driven instrument, and the energy in this system is in the moving air. And with moving air, any time you have a time-rate-of-change of flow, you have the potential for the generation of an acoustic wave. So now, let's look at this whole system from an energy point of view. For example, my speech now: I am putting maybe about a quarter of a watt into this system. Only less than one percent of that comes out as sound. So it's like I've got this tremendous reservoir of continuous energy and only a very small part of it comes out as acoustical energy. That leaves a great potential there. The opera singer stands up there on the stage at the Met singing with no microphone, with fifty or sixty pieces of orchestra in the pit, but yet that voice clearly fills that whole hall up. How do they do it with the same set of lungs and vocal chords that you and I carry around? They've learned to get that efficiency up from the order of half a percent up to seven or eight percent. That's how that gets done. But that sound generated is well above the glottis.

Goldstein:

So it's not a matter of putting more energy into the system.

Kaiser:

Right. It's not like this is a passive system and the only place where sound is generated is at the vocal chords. That's a giant crock! Now here is another thing: if you listen to somebody talk on the telephone, it only takes a second or so of conversation for you to know who is talking, in addition to what was said. If you try to do that analysis spectrum-wise, you'll find that you can't. But this approach is doing it just fine. Why? Because one's ear is looking at the modulations. It's a modulation detector. It's a transient detector. It's not simply a spectrum analyzer. It's a lot more. So the signal processing algorithms that you have to have are ones that can look not only at the amplitude modulation that's going on, but also the frequency modulation that's going on. And Herb was seeing that in his detailed air flow measurements. I hadn't realized how many experimental runs he had made. See, he'd take a hot wire anemometer and move it all around in his mouth when he was phonating—that is when he was producing the different vowel sounds—to see what was going on as far as flow patterns in the mouth. I thought maybe he'd made fifty or 100 runs, but he had done over a thousand runs! I saw the printouts and graphs of the data that he generated, and it's incredible what Herb had done! I think he had been beaten on so much by the establishment that he had just retreated into his little shell—or his big shell—and said to himself, "Look, I'm going to solve this problem once and for all so completely and get so much evidence that there's no way these fellows are going to say, "Herb, you blew it.'" I think that is just how Herb was. Herb went to MIT and his cumulative grade point average was something like 4.95, all the way through his Ph.D. At his memorial service I talked to a number of other people who saw totally different sides of Herb. And when I talked to these fellows it was obvious that we were all talking about precisely the same person. He was one person, and he had a tremendous amount of integrity.

Goldstein:

Can I go back to ask a question about something you said earlier? You referred to an organization split at Bell Labs in the acoustics department. There was DSP, where you went, and then what was the other half?

Kaiser:

It was acoustics including speech generation, vocal tract modeling, and speech recognition. It included people that also did the physiology of hearing. They even had a wet lab there for studying bats, cats, fish, and the horseshoe crab. But when the first inklings came that there was this popular movement opposed to experiments on animals, the decision was made that there would be no more wet lab work. That work just came to a grinding halt. Within about a year those researchers dispersed. Bob Capranica went up to Cornell and continued his work up there. Larry Frishkopf went back to MIT. Willem van Bergeijk went to Indiana University and set up a lab there. Those are the first three names that come to mind.

Key questions in signal processing, 1990s

Abbate:

Can you tell me what you think are the big questions, the big areas, in signal processing today?

Kaiser:

The nonlinear stuff and the time varying stuff are still very hot areas. I also want to stress again that this is an engineering agenda versus science agenda issue. It has become so easy to do so much computation using computers that people will press keys on the keyboard without thinking what they are doing. I think that there has got to be a resurgence of work on the scientific end of things. It's so easy to generate a tremendous amount of garbage that you've got to understand what it is you're doing. So it is very important that we get back to basic understanding, get a much better grounding of what science underlies the phenomenon we are looking at. We have got to be able to do that. I mean, this world is not an ideal world. It's time-varying and nonlinear. That's the first message. The next message is, the young people—or anybody, really— who are using these tools have got to understand thoroughly what assumptions underlie the tool that they are using. That will tell them what they can expect to get out. You want to know what the guts of the filter are so you can know what you have filtered out and what will pass through that filter. You've got to know that. Issues such as quantization, and, even more so, linearity versus nonlinearity. When you do linear filtering, what are you doing? How are you restricting what you can see at the output? The Teager approach to speech is just a classic example. When I talk to a few of my friends at Murray Hill, my colleagues in the speech group there, about the signal processing on speech, and I say, "Consider the speech wave." Well, the first thing they want to do is look at it through the DFT. The DFT is basically a linear tool! It's like you put on the "linear glasses," look at it, and say, "Well, I see this looks quite like a linear system." Well, that's all you can see through those tools. You've got to have a different tool to see the modulations in detail that abound in the speech wave. That was the big impetus for us working on what we call the energy algorithms, and the energy separation. For that work, the "us" included Tom Quatieri at Lincoln Laboratory, Petros Maragos (he was an assistant professor at Harvard at the time, and he's now an associate professor at Georgia Tech), and myself. Both of those fellows began seeing the things I'm talking about when they applied the energy algorithm to speech, and that got them very excited. Those two fellows became collaborators, and it really dominated Petros' work at Harvard. We three co-authored two papers that were published in the IEEE Transactions on Signal Processing in the April 1993 and October 1993 issues. It was that second big paper that received the W. R. G. Baker Award a few years ago, which is a major IEEE award for the best research paper for the year in any IEEE Transactions or the Proceedings. That made us think that, finally, people would start to take notice. Still, there are so many non-scientific issues—the turf issues, the personality issues, the where-were-you-trained issues, the don't-bother-me-with-something-I-didn't-invent-here thing. But that's the way science is, and, actually, I've been insulated from much of that.

Evolution of circuit design; computation methods

Goldstein:

Remember when you spoke earlier of how integrated circuits changed the economics of circuit design? Can you comment on how that change, and the tendency to turn to brute force computation methods as computing power grew cheaper, affected the DSP circuits that you were seeing built?

Kaiser:

It became a different ball game because you could now do things computer-wise. You could now generate a far broader class of systems than you could easily generate with the old circuit technology. That's the first thing. The second thing is kind of a subtle point, but I think active researchers should be a little more attentive to it. It is easy to change a number in a storage register, but remember that you are trying to produce signals that somehow correspond to the physical world. What you are doing mathematically may be a physical impossibility in the real world, e.g. in the reality of speech production or some other application. Oftentimes people don't realize that. It's so simple to change the numbers in a storage register to make a filter track, but what is actually generated by the algorithm that you have devised might not correspond to the real world at all. What is the equivalent in that computing machine simulation of a conservation of momentum constraint, or conservation of energy, or conservation of matter? Well, it is hardly every considered. They never think about that aspect of it. That's the engineering versus the science issue again.

Goldstein:

Was it ever the case that the easy availability of filters meant that data acquisition was done in a different way?

Kaiser:

Yes. With the capabilities you've got in the PC, if you've got any good software to design filtering algorithms, you can design real time and work with very complicated systems and generate sounds. I can generate fifty different sinusoids and add them together real time in that box. I can get very complicated signals real time. Let us get back to the music end of things. Students, young people, are into music, whether just listening, or creating it, or whatever. The algorithms to generate the music are simpler than a lot of signal processing algorithms that are used in communications systems. But it's a wonderful introduction for the student. So a lot of people have been turned on to DSP through the music and sound route. Many PCs have those pair of stereo speakers there, and a CD player and you get to hear it on your little box there. So that's going to have a considerable impact on a class of people that work on the PC.

Recent signal processing developments: chaos theory, wavelets, neural networks

Abbate:

Can you tell us about particularly important advances in the last ten years or so?

Kaiser:

Two words. Chaos theory and wavelets. There are a lot of books out on wavelets now, thirty or forty at least. Wavelets is a wonderful technique, but it's an analytic technique. It's a very efficient way to represent quite a broad range of signals, but it hasn't to my knowledge increased our physical understanding of those signals. It's another tool, but the danger is expecting that tool to increase your physical understanding of what's going on.

Abbate:

Can you tell us more about them?

Kaiser:

Wavelets are for efficient representation of a signal, whether it's one-dimensional like speech, whether it's two-dimensional like an image, where you are trying to encode that image and send a minimum amount of data over your narrow channel, down from a satellite, or what have you.

Goldstein:

You expressed your concern about some of the recent developments—people's over-reliance on computational power, for example. What's your exposure to this work that lets you know that's happening? Is it the published literature?

Kaiser:

Partly in the published literature, yes. I take theTransactions on Signal Processing and Transactions on Circuits and Systems, and I look at some of the papers in the neural network area, and I've been a SIAM member for twenty-five years. I also review papers in the Journal of the Acoustical Society of America. I look to see the kinds of things that people are doing, and signal processing now is just a lot of knob-twiddling. It's becoming much more difficult to see quantum-leap types of improvements, where you say, "This is a noticeable improvement."

Abbate:

What about the neural nets?

Kaiser:

Again, there's an issue of knowing what's going on there. Neural networks are a very powerful approach for taking essentially the output of a black box and letting the neural network decide what is the useful information. You tell it, "Well, here's the input to this box—here's all of the things that this black box is measuring—and here is what we know the result was. And here is the second test and here is the result of that one, and the third test, and so on." Then it decides, based on all those test results, a hidden network model for what's going on in there. And it works rather well. But again, it never really gives you much insight into what is really going on in that box. What is really the best way to characterize that box? If you don't care what's really going on inside, then the neural network approach is just great for taking all that data and coming up with a model that can reproduce the results. It's great for that. But again, it falls a little bit in the same category as the wavelets. It's a great tool, but it doesn't really give you a great deal of additional understanding of what's going on. Again here, the parameterization that we do in speech is what I call a front end problem. The wavelets and the neural network take whatever parameterization you give it and work with that. If we can do a better job of parameterizing than is done with the neural network, we can build a vastly improved system. So we're working on different problems. But you know, there's a tremendous amount of interest in neural networks. Any student who comes out of the signal processing area has to be well acquainted with that, and has to be acquainted with the wavelet approach, too. He or she has to understand the time frequency problem, time varying systems, and the adaptive filter. Bernie Widrow is the early worker in that area; his work dates from the 1950s.

Applications of energy algorithm to image processing, biomedicine

Abbate:

And multi-dimensional systems?

Kaiser:

Yes, the multi-dimensional systems. That's another whole area, the image processing area. That came later than the speech area simply because it takes so many more bits to represent an image than it does for ongoing speech. In television, instead of dealing with the 4 kilohertz bandwidth of speech, you are dealing with a 4.5 megahertz bandwidth picture. So the bandwidth is, literally, a thousand times greater. Perhaps that's why they say one picture is worth a thousand words. [laughter] So it's been only now that we have faster circuits, that the image processing can be done digitally and in real time.

Abbate:

Did that bring an influx of new people into the field?

Kaiser:

Yes. Just look at motion pictures like Star Wars and what LucasFilm does. With something like Jurassic Park, a lot of the special effects are done digitally. You look at the TV commercials and you see the words that stretch and twirl. It's all two-dimensional and three-dimensional DSP, the so-called geometric engines that enable you to change the projections on the fly.

Goldstein:

Perhaps if we're looking for growth in digital signal processing, it's not right to look at the technical literature at this particular stage. Maybe all the development is taking place in areas such as the ones you just mentioned.

Kaiser:

Yes. Part of my interaction with the work on image processing comes from my work in the speech area, on the energy separation. One of the classic things for me when I look at the energy algorithm is that it's simple, and it's an absolutely beautiful transient detector--an event detector, I will call it. You see, an edge is an event. It's a significant event in a picture. I was invited to U.C. Santa Barbara to give my speech talk, but when I got out there it turned out that the speech people were all somewhere else. There were only image people there at the talk. So I gave them the talk on the energy algorithm material, at three o'clock in the afternoon. Well we went out to dinner at six, and it turned out one of the students had passed his qualifiers that afternoon and so they were going to celebrate at a Chinese restaurant. On the way, one of the graduate students said, "Oh, take a look at how I've applied the energy algorithm you talked about this afternoon." I talked about it from three to four pm. He had gone back and coded it up. He was doing edge detection. We looked at his demo on the way to dinner, about an hour and three-quarters after the talk. And the algorithm was doing just as advertised. He was really enthused. Sanjit Mitra, who is a professor there, called me up about a month and a half later and said, "We have just done a paper where we are using this for image enhancement. Since we got the idea from you, may we put your name on the paper as third author?" I said, "Fine." So here those researchers were using an algorithm devised primarily for speech, and it turned out to have much more generality than that. Since that time not only has the algorithm been used in texture description and so forth in two dimensions, but also for biomedical applications. Researchers are using it now on EKGs and other signals as well. A mechanical engineer sent me a note about how he was using it trying to detect faults in a gear box, when a gear tooth is beginning to go bad. He compared half a dozen different methods, and the energy algorithm method beat every other one hands down in computational simplicity. It performed as well as the best other scheme, but with far less computation. It detected which tooth was going bad in the gear. So here is something that was developed for speech that is now being used acoustically in a gear box fault detection system. With science, it's like you're entering a giant three-dimensional cobweb. You can't move there without pulling and shaking other parts of the structure. You move around only on a small part of this big, three-dimensional web, but you do have some effect. And I think this area of research will keep me going for a good long while yet.