MIT – The Special Speech Processor, or SSP
Jack A Allweiss
After my first semester at MIT I was ready to start on my major area of study. In my signal processing class I had met a professor named Jon Allen, who was a leader in speech processing. One of the major goals for computers at the time was to have them be able to read text, and speak. The algorithms had been designed for speaking, but on general purpose computers it could not be done in real time. That means you had to crunch numbers for several minutes to get a few seconds of speech. The goal was to do this in “real time”, the machine could convert instructions to speech as fast as a normal person talked. The research was funded by the Department of Defense because they wanted real time speech feedback in areas like aircraft cockpit warning systems. The system would also have commercial applications, such as speaking appliances or as part of a reading machine for the blind (assuming the text conversion problem could also be solved).
Jon Allen had a design done by graduate students; he offered me the project of building and debugging a specialized processor called the SSP (Special Speech Processor) for this purpose. Being young an naive, I accepted the task. The task was as much management as engineering, because I would have to coordinate several undergrad assistants and departments at MIT, as well as some outside resources. Because of this I asked Allen if I could minor in Business at the Sloan school. That was a little unusual, most students that did minor work were in the same area as their major. I convinced him that in this case, it was. That turned out to be a good decision, because my work at Sloan really helped me when I founded my own companies.
My Masters Thesis Work
My second semester I took some classes to strengthen my knowledge of the algorithms the SSP was going to execute, and some additional basic courses required for my degree. The first half of the semester I worked with Jon and members of his staff who were involved in the architecture of the SSP. I had some serious questions about the feasibility of some of the design parameters, and we were able to come to final agreement on them. One of Jon’s PhD students and I then began to work on the detail design of the SSP.
The second half of the semester was taken up primarily with the grunt work of generating the schematics of the SSP. In those days schematics were mostly drawn by hand, and that is what we did. The SSP had a lot of common logic, so it was not too bad. Next, we had to decide how to build it. Jon and the PhD student wanted to just have a bunch of undergrads wirewrap the SSP from the schematics. Wirewrap was a method where each IC is placed in a socket on a printed circuit board, and then wires are strung from pin to pin. I objected to the plan. It was clear to me from my brief experience at Burroughs the previous summer that wiring the SSP by hand would not work, it would be full of errors and difficult to debug. At Burroughs automatic machines, fed from a tape, wired the circuit boards. The tapes were generated by transcribing the schematic to punch cards using a type of circuit description language that a computer program translated into wiring instructions. The problem was that MIT did not have such a program, but I knew who did.
At Wayne State I had worked on two projects, the ultrasound and kidney function, and for both I used a program on the WSU mainframe that was originally developed at University of Michigan. I contacted the U of M and asked if I could have the program. They agreed to supply MIT with the program at no cost. About two weeks later six boxes of punch cards arrived. I still had a problem. The SSP was to be built in a new integrated circuit family called ECL (emitter coupled logic) because of the speed required to do real time speech. The Wirewrap program from U of M was designed for TTL, an older logic family. Because of the way ECL worked, special rules had to be followed in connecting the circuits. To use the Wirewrap program I would have to modify the wiring algorithm which was the heart of the program. I met with a couple of mathematics professors who were experts in topology, and working with them I came up with an algorithm that would work. I coded it and tested it on a small sample design. It worked! By now the second semester had ended, and we were going into the summer months.
During the second semester one of the business classes I took was on systems management, how to organize large projects. I put what I learned to work right away. I created a detailed flow diagram of what had to be done to finish the SSP. I was slated to Graduate in January of 1976, and I wanted to make that schedule. Based on my flow chart, it would take until May to get the SSP up and running. I reviewed the schedule with Jon Allen. He was impressed, none of his students had taken classes at Sloan, and so the tools I used to schedule the project were new to him. I told him I would try to get the SSP finished by January, but I could not promise that. The SSP was not a stand alone computer. It had to be interfaced to a DEC PDP-9 computer to load its instructions and data, and unload its results. Jon agreed that if I got the basic SSP done and debugged (i.e. running its stand alone diagnostics) that he would consider it finished, and he would have another graduate student finish the DEC interface and software. That seemed reasonable to me, and we agreed.
I had two undergrad students transcribing our schematics to punch cards, while I began coordinating all the parts needed to build this machine. I had to work with MIT purchasing to buy all the ECL logic chips, the boards and power supplies. I worked with the MIT in house fabrication group to build a chassis to hold the SSP and a cabinet to house the whole thing. The SSP needed a front panel for controls, and that had to be designed and fabricated. To wire the printed circuit boards I had to find a company that offered that service. I found WrapCon in Burlington Mass. Bob Lee, the sales manager came by and agreed to take on the job (ECL was new, and required some special handling). That summer was a flurry of activity. In addition I was taking two classes I needed, one business and one engineering.
Working on the SSP was not only my thesis project, it was my Research Assistantship. Burroughs only paid for my tuition. I had to pay for my books and living. Of course Patty’s salary covered most of that expense; it was still very tight so the RA money came in handy. About mid summer Jon Allen came to me and said that Lincoln Labs was building a new processor out of ECL but did not have the tools to wirewrap them. Our contract was partially funded by the Navy and the last review meeting someone had mentioned the problem Lincoln Labs was having. He asked if I wanted to work for them to solve the problem. Given that they were going to pay me some additional money, I agreed.
The next day I went out to Hanscomb Field in Lexington where the project was based. The project manager talked with me for about a half hour, and hired me on the spot, provided I pass the “secret security clearance”! Well, it took them about a week, but I get a call telling me to report to work at Hanscomb. I also got a call from family and friends asking why the FBI was calling them about me! I told them I was working on a secret project, but I could not tell them what it was, (the old joke, if I tell you I would have to kill you!). It turns out that the processor was part of a missile defense upgrade, but that is all I knew (you know, “need to know” basis). My main job was to get the Wirewrap program running on the mainframe, and consult with them if there were any routing problems, as I wrote the algorithms. There were a few problems because their design stressed the algorithm more that the SSP did. I ended up going out to Hanscomb two or three times a week for about three months. It was good money. Once I was on the base and needed to leave to drive back to MIT for class (I usually took the T to MIT, but on days I was going to Hanscomb I had to drive). I am about to leave the base when the siren sounds, the base is on lockdown, no one could come or go. I talked to the MP, “hey I’ve got class in an hour”. He told me to just cool my heals, sometimes lockdowns last five minutes, some times five hours! Lucky for me this one lasted about 15 minutes and I was able to get to class on time.
So in late summer, early fall of 1975 I am 1) Working for Jon Allen on the SSP, 2) Working for Lincoln Labs on the secret project, 3) Going to class, and 4) Starting to write my thesis! Yes, I was busy. Even so, Patty and I took some time for fun. I learned to sail on the Charles River that summer, and Patty and I would go out on the sale boat sometimes. It was fun sailing in the city. Our financial situation was better now, so we could afford to go out for dinner in the Italian North End, or on Beacon Street in back bay were we loved to have the crapes. During this summer we took trips to Newport Road Island with friends and family. We enjoyed the lobster at Cape Ann, and took to cog railway to the top of Mt. Washington.
By the fall I needed to begin to think about going to work. I briefly thought about a PhD, and Jon Allen tried to convince me it would be a good idea, but a PhD was really only valuable if you were going to teach, and I wanted to go into industry. At that time PhD’s did not command much more than Masters in the business world. Although my course work was done by now for my major in Electrical Engineering, I agreed to take two more courses so I would qualify for the EE degree, basically a PhD without the thesis. I could them come back and work on a thesis without having to do any additional course work. I also took the “Prelims”, a set of tests you needed to pass to be accepted into the PhD program, and passed them. I also took my last minor class at Sloan on Managerial Finance, which was very valuable later in my career, so I was taking three courses my final semester where most students were taking only one or two and working on their thesis.
In early fall the SSP was behind schedule. The MIT mechanical department was behind in getting the chassis done. The power supplies were running late. We discovered some logic errors after the circuit boards were already wired, so we had to have them reworked. These are all normal things in the design of a complex computer system, but being totally green I did not know that. In fact I found out that rarely were things designed by the academics actually built, I guess I was the only one either stupid enough or bold enough to try! At this time I also began to reconnect with Burroughs. They set up interviews for me at several of the plants. The Tredyffrin Plant in Pennsylvania built the largest Burroughs computer at the time, the B7700, was the primary plant they were steering me towards. I also interviewed at the Memory Systems plan in Piscataway New Jersey.
In the latter half of the 70’s there was an economic slowdown, these were the Jimmy Carter days of “may lays” and people were not hiring, so starting salaries were not super. I did not restrict my interest to Burroughs. Both DEC and Data General had close ties to MIT, I interviewed at both. Data General flew me down to North Carolina where they were starting a new group to design their 32 bit minicomputer. That design effort was the subject of the book “Soul of a New Machine” which gave a great account of what it takes to get a new computer system launched. I was tempted to take the DG offer, but in the end declined.
About this time we had a surprise visit from Ken Olsen, the legendary CEO of Digital Equipment Corporation (DEC), at that time the leader in minicomputer systems. At the time many thought that DEC might eventually challenge IBM for the leadership of the computer industry. While DEC was smaller than Burroughs, it was growing faster. It had begun to build systems that would be considered entry level mainframes (the DEC system 20 and VAX). It is ironic that by the early 90’s DEC went into chapter 11, and was then bought by Compaq Computer. But for now, DEC was a leader, and Ken Olsen was not a humble guy. Jon Allen was showing him around the lab, and our SSP, in a partial state of assembly, was shown to him. I will always remember Olsen’s reaction, “You should not be doing this, leave it to the computer companies to build the systems, you just come up with the design ideas”. I took his comments personally, and was hurt by them. In reality, it had nothing to do with me and everything to do with how Ken Olsen viewed the world, a view that would eventually lead to the downfall of his company. In retrospect, he was partially right. MIT, really did not have the resources or discipline to manufacture something, they should have reviewed the design of something like the SSP with DEC or some other company. In fact that is a model that is used today. Universities design algorithms or systems, patent them, and then license the patents to commercial enterprises to actually build, but that model did not exist in the mid seventies.
Finally, near the end of the semester the SSP was ready for debug. I was finishing my classes and my thesis, so getting time to debug the system was difficult. We fired up the system and immediately we had problems. The system did not have a stable initial state, something that Jon and the other PhD students forgot to consider. I began making modifications, but we did not have a full time staff to make changes, so it was slow going. Just as the term came to an end I got the system to power up in a stable state.
About this time I also made my decision on what job I was going to take, and it was a surprise to Burroughs, instead of the large system plant, I picked Piscataway, the Memory System plant. My experience with the SSP told me that memory was going to be a key component of systems in the future, and I wanted to learn more. Also, at Piscataway I would be able to travel to all the systems plants because at that time Piscataway supplied memory to all the plants. Finally I would be a big fish at Piscataway, and not so big a fish at Tredyffrin.
The end of the semester came, and the SSP was still not running any diagnostics. I had a deal with Jon, and I wanted to keep my end of it. This was a problem, because our apartment lease was up at the end of January. Patty got a transfer with her employer John Blair and was going to start work in their Fifth Avenue New York headquarters on February 3rd. We had an apartment lease that started February 1st. I met with Jon, I told him that I thought I could get the diagnostic running by the middle of February, and I would stay until then, delaying my start at Burroughs by two weeks.
This caused a lot of hardship for Patty. Since I was not working for Burroughs yet, we had to move ourselves to New Jersey. She would have to move down to New Jersey without me for about three weeks. To make matters worse the move was more expensive that we thought, and we had no money. We rented a U-Haul, packed up our stuff, and moved down to New Jersey the last weekend of January. We unpacked, returned the truck, and I drove back up to MIT. The next day Patty was on the train to New York. That first day Patty went to New York she did not have a dime, she asked her boss for an advance and that tided her over
Meanwhile, we did not have an apartment in Boston anymore, so I was living out of my office at MIT. I ate at the cafeteria and when I did sleep, it was on a couch in my office. I showered in the gym. But mostly (about 18 hours a day) I worked on the SSP. I finally got the system to execute its basic instruction loop. The biggest part of the machine was a hardware multiplier custom tailored to speech processing. I was able to get that part to work, but there were fundamental problems with the interface logic. I was out of time. I meet with Jon and his PhD student and reviewed the status of the SSP. The signal processing instructions worked but there was no way to test the interface, as that was not built yet. I looked at that logic and from my experience with the rest of the machine I saw that major timing problems existed in the interface. I recommended that the interface be redesigned.
One thing I learned that would serve me well was that design controls were essential when debugging complex systems. Early on in the SSP debug I got hopelessly lost in the design. I had made changes that I had not documented properly, and this compounded other documentation errors. Finally, I had to stop for a full day just to find out where I really was. I did not understand at the beginning how design change control was so important in a large system debug.
I left MIT the middle of February.
In the end, the SSP could do most of its diagnostics, and one of the older technicians who worked for RLE and assisted me was very interested in the design and fixed a few instructions that I failed to complete. Jon Allen assigned another graduate student to complete the interface to the DEC system, but he never did. The PhD student was near the end of his PhD thesis and lost interest in the SSP, so from that perspective the machine was a failure, it never really worked. From another perspective it was a success. Just running the diagnostic, after I made several design changes during debug, proved such a design could work. About nine months later another graduate student took the SSP design and decided to implement it using a new technology that was taking the industry by storm, VLSI CMOS.
MIT was a leader in new design methodology required to implement CMOS, and the VLSI group was looking for new applications to show off the technology. The SSP, with its very regular and modular design, was a perfect candidate. About a year after I left, Jon Allen’s team working with the CMOS group developed a working version of the SSP on a chip. That core design is the basis of almost every real time (not stored) speech synthesis system on the market today. How many of those are there? Millions. Reading machines for the blind, automobile GPS, airline cockpit systems, talking toys, any time a computer talks with an unlimited vocabulary, the SSP is there, buried deep inside.
With a lot of new book knowledge, and an unbelievable amount of practical knowledge from the SSP experience, I was off to make my way in the real world at Burroughs; I had just turned 23 years old.