Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
200 MICROPROGRAMMING All digital computers consist of assemblies of hardware which either store, route, or process (combine algebraically) bits of data. In the simplest computer, the assemblies are simply wired together in a permanent manner; the computer then performs one unchangeable operation whenever it is pre- sented with input bits. Its internal "logic" is mechanized by hard-wire connections; its only "control" is the on-off switch. Its "instruction set" simply describes its one operation. It is inflexible but, for some applications, may be faster and cheaper than any competitor. On the other hand, its marketability might be very limited! The next most complex computer would have some toggle switches on it which could permit the user to change the "hardwire configuration" in ways which would affect signal storage, routing or processing. Several "logics" are now possible and the instruction set is more complex since it must describe what happens for all possible positions of all the switches. The "control" is now a full set of switches. Obviously, by changing switch positions during the opera- tion, one could perform a whole time sequence of operations on the original data bits. The switches might control any- thing from the smallest piece part in the computer to a whole major assembly. Even at this primitive level of com- puter we can see the tradeoffs that must be made by the man- ufacturer as he tries to produce a marketable product: l. Too few switches and the applications are too few. 2. Too many switches and the computer becomes expensive, the wiring more and more complex, and the probability of successful operation under all switch condi- tions becomes marginal. Furthermore, if the manufacturer intends to produce better and better computers in the future, the more switches and possibilities provided to the user, the harder it is to guarantee that the next computer in the series can still do the job of the present one, especially if the old "instruction set" which says which switch does what is to be retained. Users who have invested major amounts of time and money in figuring out elaborate sequences of switch positions ("software") are going to be particularly concerned with preserving this investment as long as possible. As a result, the manufacturer comes up with a configura- tion of hardware and switches (an "architecture") which he
201 hopes has enough applications to be a marketable product. Most users will be satisfied if the manufacturer has done his job well. There will always be some users who will find the architecture cumbersome and inefficient for their purposes and for whom a slightly different set of switches and instruction sets might have made a great difference. The special problems of these users are often solved by adding other special purpose equipment ("associative processing"); a typical example is a simulation of aircraft performance using actual aircraft hardware as part of the computer-assisted setup. There is also the alternate possibility of the user physically modifying the manufacturer's hardware (architecture), but clearly this could lead to trouble unless very skillfully done. At the very least, the manu- facturer's "warranty" is probably voided in that the user can't reasonably expect the manufacturer to guarantee that the computer will still work properly, nor to guarantee compatibility with other computers in the manufacturer's line. However, the user may be much better off if the modi- fication is well done, and if he has no concern over compat- ibility with future machines. With this background it is fairly easy to see the im- pact of very new technology on "microprogramming". Micro- programming as a word traces back to the early l950's. Its purpose, then and until recently, was to try to organize and discipline the process of deciding on computer architec- ture. As such, microprogramming was generally an aid to the computer designer rather than the user. As might be expected from the foregoing discussion, manufacturers were reluctant to have the users tamper with something as basic as computer architecture. But then, within the last few years, it became possible to build "chips" containing hundreds to thousands of computer elements each. The nature of the technology led to extremely high reliability per chip. But clearly a chip of so many elements which was "hard-wired" to perform only one function would be too inflexible for general use. This problem was solved by having a number of connections to the chip (typically dozens), equivalent to switch connec- tions, by which the same chip could be made to perform differ- ent functions. One obvious advantage of such chips is that a computer can now be made with many identical chips, with the external wiring to their connections determining what function they each perform. Another straightforward improvement is to have the external wiring changeable by using prewired plug boards. A step beyond that is to use computer cards rather than plugboards. Conceptually, the wiring, the plug boards and the computer cards are all forms of "fixed memory" specifying the "control" or "architecture" of the machine.
202 Enter now another new technology: cheap, reliable, and fixed memory and the possibilities for creating many differ- ent architectures abound. Indeed, computers built around such technology can be made to look like (emulate) other computers' architectures, thus helping to break the overly- tight relationship between specific computer hardware, the instruction sets, and the user's software. Conceptually, one simply tells his own (host) computer, by changing the fixed memory which controls its architecture, to "become" another (target) computer for which one has soft- ware already available. In practice, hardware details can still cause trouble, particularly if the host and the target computers are quite dissimilar at the hardware level; thus, if one replaces his own computer by a new model, the fixed memory controls developed for the first computer are unlikely to be directly applicable to the new model. Typical problems that arise are those involving system synchronization, time delays, word length, techniques for handling data overflow or discard, etc. And finally, one can replace the fixed (read-only) memory of the architecture by one which can be electronically changed prior to, or even during the solving of a problem on the computer. One can thus literally change computer archi- tectures in midstream of a problem, taking advantage of the best properties of different architectures as one goes along. This whole new world of technology is usually labeled "microprogramming". As computers and telecommunications become increasingly linked into very fast, high capacity automated systems, microprogramming will help to enable these systems to operate with greater efficiency. Heretofore, the computer and tele- communications technologies have tended to go their separate ways and have not been tied sufficiently closely to bring about the full range and flexibility of their potential as parts of a combined system. As computer communications net- works grow in size and become international in nature such cpmpatibility will be essential to their effective function- ing. I believe one can now draw the following management conclusions about microprogramming: l. In the past, the principal direct benefactor of microprogramming was the manufacturer, with the user benefit- ing indirectly through the better computers being offered through the years. It is now possible for the user to take direct advantage of microprogramming as well by changing the
203 basic architecture of his computer; however, at present, the user must be a relatively sophisticated one (e.g., a Federal agency, a major industrial firm, etc.). 2. A new division of labor in the computer business is predictable, with the "chip" manufacturers increasingly dominating the hardware end and the computer suppliers increasingly concentrating on producing "different computers," including those of their competitors, by micro- programming different architectures. 3. The United States is unlikely to fall behind in this area. It is presently far ahead in the hardware tech- nology and in the understanding of the techniques of micro- programming. It is attacking with some vigor new problems such as multi-level security with one machine, generation of software prior to the commitment to hardware, efficient processing of radar and acoustic signals, and application of computers to avionics. Supplemental research funding by the NSF would appear unnecessary. On the other hand., a close watch on this field by the Department of Commerce should be encouraged since the United States has no monopoly on inexpensive solid state technology,or on the mathematics of computer design and use. A combination of good decisions by foreign competitors and unfortunate ones by US firms, particularly in "chip" design and read/write storage, could lead to a serious loss of market for periods as long as 3-4 years. E. Rechtin Department of Defense (Telecommunications)
204 SPEECH ANALYSIS Speech analysis is the process which extracts the information-bearing components of speech and converts these into some form of a code. The elements of speech sounds are phonemes, syblets, syllables and words. A phoneme is the smallest speech unit that serves to dis- tinguish one utterance from another. A syllable consists of an uninterrupted unit of utterance and contains one or more phonemes. However, even in what is considered to be a syllable, there may be some voiced and unvoiced inter- ruptions. A syblet is a part of a syllable bounded by such interruptions. A word is the smallest unit of speech that has meaning when taken by itself and consists of one or more syllables. The objective of speech analysis is to recognize speech elements. Therefore, the number in each category becomes important because the complexity of the analyzing system is a function of the number of speech elements involved. Forty phonemes, 800 syblets, l400 syllables and l0,000 words con- stitute the major part of the American language. In the selection of speech elements for analysis, any subdivision of a word or syllable which a machine can carry out and classify reduces the total number of speech elements to be recognized. At the present time, the speech elements which appear to offer the greatest promise for analysis are an optimized selection of syllables, syblets, or phonemes according to whichever can be segmented and analyzed. In order to analyze the different speech sounds there must be some means for the segmentation of the flow of speech. The segmentation involves words, syllables, syblets and phonemes. A boundary occurs when the sound amplitude drops to zero, or near zero, and the transition takes place from voiced to unvoiced sounds. In general, the vowels constitute the voiced sounds, and the consonants, the unvoiced sounds. A sound wave may be completely described in terms of amplitude and frequency of the components and the time. Therefore, the fundamentals of any analyzing system must be based in some manner upon amplitude, frequency and time. The real trick in speech analysis is to employ the three funda- mental parameters in such a manner that the speech elements can be differentiated along unique features, characteristics and procedures which lead to the ultimate objective, namely, recognition.
205 The main features, characteristics and processing required for the recognition of speech elements are shown in Table I on page 207. The first process is segmentation by amplitude pauses, voiced and unvoiced, and voiced and unvoiced transitions. In the voiced elements there are initial vowel and non-vowel sounds. The speech sounds are normalized with respect to ampli- tude to compensate for difference in loudness, and with respect to time intervals when there are no significant changes in speech sounds. The envelope, spectral distribution and time as depicted in Table I provide specific information on the recognition of the speech sounds. Quantization and data reduction are employed for further classification of the speech sounds to facilitate recognition. The instrumentation required to carry out the processes depicted in Table I is indeed quite complex and beyond the scope of this brief summary. There are two types of speech input operations, namely, a cooperative speaker, where a person is willing to speak a set of words so that the system will adjust to dialect, etc., and the uncooperative speaker where the person is not required to say any words. In the cooperative type of operation, an accuracy of 98% or better can be achieved. For an uncoopera- tive speaker, the accuracy may be quite high for some speakers and very low for others. The speech elements are converted into a code. The code may be used for the transmission of speech by the use of a speech synthesizer, the input to a computer, the control of machine operations, etc. At the present state of the art, the applications must be restricted to limited vocabularies and cooperative speakers. If the speaking rate is 90 words per minute, the band- width required for the transmission of speech employing l0,000 words is depicted in Table II. For monosyllabic words, the speaking rate is l50 per minute. The number of syblets is twice that of words. The approximate number of phonemes is four times that of words. The frequency bandwidth for the conventional analog transmission of speech is 3000 Hz. This corresponds to a bit
206 rate of 30,000 bits per second for a signal to noise ratio of 30 dB. Referring to Table II, it will be seen that there is a tremendous saving in bandwidth by the use of speech elements. The data of Table II is for a signal to noise ratio of 30 dB. The data of Table II represents only the recognition aspect of speech and does not include accent, inflection, tone, etc., and artistic properties of speech which would increase the bit rate. The importance of speech analysis for speech trans- mission, input to computers, control of machines and other applications provides the reason for the extensive research programs. Harry F. Olson RCA Princeton, N.J.
207 PROCESS FOR RECOGNITION OF SPEECH SOUNDS ( Amplitude pauses The main features, characteristics and processing required for the recognition of speech sounds Segmentation Normalization Envelope Spectral Distribution Amplitude Distribution Time Quantization Data Reduction (Voiced-unvoiced-voiced Unvoiced-voiced-unvoiced / Amplitude / Absolute Relative Timeâno change Growth Duration-steady state Decay Bilateral unbalance Absolute Relative Pattern Absolute Relative ( Absolute \ Relative I Epochs ( Frequency < Amplitude I Time Compilation Manipulation Logic applications Programming Table I
208 TRANSMISSION CHARACTERISTICS OF SPEECH ELEMENTS Number of Transmission Transmission speech of speech rate. Transmission Speech elements elements per bits per bandwidth. elements required second second in Hertz Words 10,000 1.5 21 2.1 Syllables 1,400 2.5 27 2.7 Syblets 800 3.0 30 3.0 Phonemes 40 6.0 36 3.6 Table II
209 SPEECH SYNTHESIS Speech synthesis is the process of producing speech from some sort of a code. There are many forms of speech synthesizers as, for example, an electrical analog of the vocal tract, spectrum reconstruction techniques, computer simulation and recorded speech elements. The analog of the vocal tract consists of electronic frequency and hiss generators coupled to electrical reso- nators. The generators simulate the glottal source and noise sources. The electrical resonators simulate the mouth and nose cavities. The electronic generators and the electrical resonators are controlled by the coded in- put. Very intelligible speech can be obtained from vocal tract analog systems. The spectrum reconstruction speech synthesizer, as the term implies, constructs the spectrum of the original speech from a code input obtained from the analysis of the original speech. The first development was the Voder con- sisting of a buzz source and a noise source connected to l0 contiguous band pass filters. The frequency and filters were controlled to produce the speech sounds. The Voder can also be operated from a speech analyzer and thereby produce speech transmission. These systems are termed Vocoders. Many types of Vocoders have been developed for the transmission of speech. The approximations of vocal transmission can be repre- sented by linear differential equations. Such equations can be approximated by difference equations which can be pro- grammed in a digital computer as arithmetic operations upon discrete values of the variables. In this manner, speech can be produced by a digital computer. The speech elements such as words, syllables, syblets and phonemes, can be recorded on tracks on a magnetic drum and reproduced by means of a track selector. A decoder actuates the track selector from the coded input. This type of speech synthesizer has been in use for many years, calling out the floors in an elevator, giving the time by telephone, providing speech output from computers, etc. There are many versions. In some, words are used, in others, phonemes and syllables are combined to form the words.
2l0 From the standpoint of performance and applications, speech synthesis is more advanced than speech analysis. Harry F. Olson RCA Princeton, N.J.
