SYNTHESIS' OF STOPS, FRICATIVES, LIQUIDS AND VOWELS BY A COMPUTER CONTROLLED ELECTRONIC VOCAL TRACT ANALOG. ' b y KENNETH A.

Transcription

1 SYNTHESIS' OF STOPS, FRICATIVES, LIQUIDS AND VOWELS BY A COMPUTER CONTROLLED ELECTRONIC VOCAL TRACT ANALOG ' b y KENNETH A. SPENCER B.A.Sc, University of British Columbia, 1967 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in the Department of Electrical Engineering We accept this thesis as conforming to the required standard Research Supervisor Members of the Committee.» Head of the Department Department of Electrical Engineering THE UNIVERSITY OF BRITISH COLUMBIA December, 19 71

2 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the Head of my Department or by his representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of Electrical Engineering The University of British Columbia Vancouver 8, Canada Date December 21, 1971

3 ABSTRACT The speech synthesizer described in this thesis is a computercontrolled solid-state analog of the human articulatory system. The synthesizer's design overcomes two serious difficulties which have previously hampered vocal tract analog realization; namely, the realization of time-varying inductors and capacitors which require a minimum of computer service and whose important parameters are relatively insensitive to l.arge variations in terminal voltage and current, signal frequency, and instantaneous element value; and the realization of a control scheme which requires a minimum of computer storage capacity and service to the synthesizer. Present software provides control for the articulatory analog, allows on-line construction, evaluation, and modification'of words, displays articulatory parameters during speech production, and provides for automatic checkout of synthesizer hardware. The analog was used to synthesize vowels, semi-vowels, frica tives, stops, and English words. Spectrograms of synthesized phonemes were compared to published data. Articulator positions and timing parameters for synthesized phoneme sequences and English words were evaluated using subjective listening tests. Fifty English^Xtfqrds were synthesized to demonstrate intelligibility, structure and complexity of synthesis rules, and to demonstrate that the synthesizer's accent can be learned quickly.

4 TABLE OF CONTENTS Page ABSTRACT i i TABLE OF CONTENTS i i i LIST OF FIGURES LIST OF TABLES.... ACKNOWLEDGEMENT... v vii viii 1. INTRODUCTION 1.1 Uses of Speech. Synthesizers Past Work Scope of the Thesis ELECTRONIC SIMULATION OF THE HUMAN ARTICULATORY PROCESS 2.1 Description of the Vocal Mechanism How Sounds are Produced An Electrical Analog of the Human Vocal Tract Overall Description of the Vocal Tract Analog HARDWARE AND SOFTWARE 3.1 Time-Varying Conductor Time-Varying Inductors Time-Varying Capacitors., Digital Interpolator Glottal Source Noise Source Control Overall Hardware Synthesizer Software. 41 Supervisor 41 Display and Light Pen 43 Teletype. 43 Vocal Tract Analog 43 A-D Handler Sequence of Events During Synthesis of a Word TESTING THE VOCAL TRACT ANALOG 4.1 Introduction to the Vowel Tests Description of the Vowel Listening Test 49 i... IX x

5 Page 4.3 Results of the Vowel Listening Test Frequency Response and Formant Frequency Measurements Formant Bandwidths of the Vowels Introduction to the Fricatives Listening Test for the Fricatives Results of the Fricative Test Frequency Response of the Fricative Configurations SYNTHESIS OF THE STOPS 5.1 Introduction to the Stops Methods of Closing the Vocal Tract Target Configurations for the Stops Methods of Producing initial stops Stop-Vowel Test I Results of Stop-Vowel Test I Fricative-Stop-Vowel Tests Results of Fricative-Stop-Vowel Tests Stop-Vowel Test II 5.10 Results of Stop-Vowel Test II Spectrographic Study of the Voiced Stops Summary of the Stop Synthesis Results THE ENGLISH WORD TEST 6.1 Introduction to the English Word Tests Synthesis Rules for the English Words The English Word Tests 6.4 Results of the English Word Tests I CONCLUDING REMARKS 7.1 Realization, Cost and Capability of a Computer Controlled Vocal Tract Analog 7.2 Further Work Summary HO APPENDIX PHONETIC TRANSCRIPTION SYSTEM H 1 BIBLIOGRAPHY I 1 2 iv

6 LIST OF FIGURES Figure Page 2.1 (a) Midsaggital section of the human vocal mechanism 9 (b) Functional diagram of the human vocal mechanism Z2 Vocal tract profiles for the principal phonemes in the English Language (a) Right cylinder (b) Electrical analogs for a short right cylinder 14 "2.4 Schematic of MINERVA (a) Discrete-state time-varying conductance and its control circuitry 19 (b) Stepwise approximation G(t) to continuous time function G(t) Circuit of Riordan for simulating a time-invariant inductor 3.3 (a) Ungrounded time-varying inductor (b) Series combination of the k resistor and switch in. G(t) 3.4 (a) Q of steady state inductor plotted against control number B and frequency f (b) Steady state inductance for the circuit of Fig. 3.3 vs. frequency f and control number B (c) Same as (b) for L=1.0H Circuit used to simulate capacitor Final realization of time-varying capacitors Input capacitance C and quality factor Q vs. variable resistance R^ for capacitor in Fig Input capacitance C and quality factor Q vs. binary control number B for capacitor in Fig Quality factor Q vs. frequency for capacitor in Fig. 3.6 for various values of C^ and R^ Control number N as a function of time using up-down counter, comparator and variable rate clock (a) Glottal excitation source (b) Waveforms ^ 3.12 Spectrum of voicing source used and spectrum of source used by Stevens et al.. 33

7 Figure Page 3.13 Variable rate pulse generator used to simulate changing inflection at voicing source Variable rate clock pulse produced by circuit in Fig Noise generator used to simulate friction and aspiration Illustrating modulation of the noise source by the glottal source * Digital comparator and clock memory matrix Overall hardware facility Function of program blocks Timing information and source amplitudes for the word LEASH Timing used for vowel test Data used in vowel tests Frequency responses of vocal tract analog for 10 selected vowel configurations First formant-second formant (F^-F^) data Formant frequency of vowel configurations from Stevens et al First formant bandwidth plotted against first formant frequency (a) Vocal tract configurations for synthesizing fricatives... (b) Frequency response for configurations in (a) Timing controls used in fricative-vowel tests Timing control for /ho/ and /ha/ in fricative-vowel test Timing used in pilot fricative-vowel test CVC Configurations used to simulate the stops Control timing used in stop-vowel test Control timing used in fricative-stop-vowel test Synthetic spectrograms of stop-vowel combinations 95 vi

8 LIST OF TABLES Table Page 4.1 Results of vowel listening test Signal-to-noise ratio for 10 selected vowel configurations used in Fig Bandwidths of vowels measured by Dunn >.and bandwidths of vowels synthesized on MINERVA Results of vowel-fric.itive test Results of pilot test with fricative-vowel combinations CVC Results of stop-vowel test I S3 5.2 Results of fricative-stop-vowel test Results of stop-vowel test II The maximum response rate for each stop with various vowels Results of English word test Percentage scores in English word test for each word Incorrect responses for each word and percentage of time each was made during tests 105 vii

9 ACKNOWLEDGEMENT I wish to acknowledge the assistance I received during the years of work on my thesis. Dr. Robert W. Donaldson, who supervised the thesis and provided encouragement throughout the project. Dr. Andre-Pierre Benguerel, who heped me through the phonetics and provided subjects for listening tests. Jean-Pierre Steger, who quickly and efficiently built MINERVA. Gordon McConnell, who did the maintenance and changes for MINERVA. Linda Morris, who quickly typed the thesis. The National Research Council who supported me and the project. My wife, Sharon, who gave encouraging support to me throughout the thesis, typed the first draft and spent many hours proof reading.

10 1 1.1 USES OF SPEECH SYNTHESIZERS 1. INTRODUCTION Speech synthesis has been studied since the late eighteenth 1 2 century, when Kratzenstein and Von Kempelen made mechanical talking machines. Potential uses of speech synthesizers today are many-fold. One of the main uses would be as an output device for a computer. If computers could speak their answers as well as print them and display them graphically, the use of digital computers could be broadened. The present cost of putting computer terminals in homes and offices is s t i l l relatively high, but if a terminal consisted of the telephone, then vast amounts of information could be available to virtually anyone having access to a telephone. Other typical uses of speech synthesizers are; reading machines for the blind, teaching machines, communication with the computer in spacecraft and aircraft without a bulky terminal or console, and automatic information services. Another important use of speech synthesizers 3 is in speech research. As Mattingly points out, when developing a theory for language, the phonetician can never be sure that the utterances he receives from speakers have followed the rules that he has set down. Having a speech synthesizer can help in learning about the speech process. The closer in concept this synthesizer is to the human vocal tract the more one might learn about how humans produce speech. 4 5 As'Holmes et al in 1964 and Flanagan et al in 1970 pointed out, synthesized speech need not be identical to a natural utterance for many applications. Recognition as a possible utterance in the dialect is sufficient. The machine can be permitted to have a machine accent and s t i l l be useful. i

11 2 1.2 PAST WORK Speech synthesizers fall broadly into two classes, terminal analogs and vocal tract analogs. Each type can be realized physically using hardware, or can be simulated on a general purpose digital computer. A terminal analog attempts to reproduce the spectral properties of the speech being synthesized. Generators of various frequencies may be used to give a certain spectrum, or a source rich in harmonics may be filtered. The filters may be serial, or in parallel. The spectrum of speech contains prominent peaks, called formants. Reasonably intelligible speech can be synthesized by using three bandpass filters to reproduce the first three formants, although terminal analogs have difficulty reproducing speech where the spectral characteristics change rapidly. A complete study of both terminal analogs and vocal tract analogs was given by Flanagan in One of the earliest terminal analogs was the pattern playback device developed at Haskins Speech Laboratory in The control was painted on a belt, and this control intensity modulated harmonically related frequencies as the belt ran over photocells. Since this early attempt many terminal analogs have been developed, including PAT 7, OVE I 8, and OVE II 9, SPASS 10, and Glace Holmes. Some terminal analogs such as OVE III and one developed by 13 Cooper and Epstein are computer controlled, giving them a great deal more flexibility. A vocal tract analog attempts to model the equations for pressure and volume velocity of the air in the human vocal tract during speech production. The model usually assumes that the human vocal tract can be modelled by a series of right circular cylinders of various areas.

12 In 1950 Dunn made a static analog with twenty-five sections, each 2 modelling'.5 cm of length and 6 cm of area. A constriction corresponding to the tongue hump could be inserted, dividing the vocal tract into two cavities. By moving the position where the tongue was inserted different vowels were produced. Stevens et al"^ in 1953 made a thirtyfive section analog, with each section simulating.5 cm of length. 2 2 The area of each section could be varied from.17 cm to 17 cm by means of a bank of switches. consonants /f, s,.[/*. The analog produced the vowels and the fricative Both the above analogs were static in that they could only produce one sound continuously. To produce another sound the bank of switches had to be changed manually. In 1956 Rosen"*"^ built a fifteen section analog that could store two configurations. Changes from one configuration to another were made by a specially designed control system. This vocal tract analog produced fricative-vowel and vowel-fricative combinations. Hecker in 1962 added a nasal analog to this vocal tract analog and produced nasals. 18 In 1962 Kelly and Lochbaum developed a program that simulated the vocal tract. The output was a digital tape that represented the sound pressure waveform. This program took about forty seconds to 19 produce one second of speech. Matsui et al have built a vocal tract analog that runs from a computer. The compu-ter calculates the parameters for speech every five msecs and generates a control tape. The control tape then runs a vocal tract analog off-line. This synthesizer is capable of producing short stories. A IPA or modified symbols are used throughout the thesis and are explained in Appendix A and Fig. 2.2.

13 20 Rice in 1970 developed a software analog. In this analog one configuration is stored in a computer for every twelve msec of speech. The excitation of the vocal tract is also stored. From the configuration and the input, the program calculates the output. At present, this program produces vowels and other voiced sounds. There are plans to extend it to include sounds that use other excitation. This program takes one hundred seconds to calculate one second of speech. Not all synthesizers are distinctly terminal or vocal tract analogs. One of the synthesizers used by Flanagan et al^ is a mixture of the two. The starting point is a vocal tract configuration and the computer solves for eigenfrequencies of the configuration every ten msecs. This solution information is then used to run a terminal analog. The problem of automatic speech synthesis can be divided into two main areas. One is the development of higher level rules to produce sentences from phonemic inputs. Rabiner and his co-workers 5 ', Mattingly and Tatham are among the workers in this area. The other problem is to produce better synthesizers that lend themselves to being controlled by simple rules. 1.3 SCOPE OF THE THESIS The work presented here is a continuation of Rosen's work. Rosen's analog had some serious drawbacks. Because it was made of vacuum tubes, the parameters drifted so much that the analog required frequent adjustment. The inductors used were saturable reactors which required a non-linear control. The value of the inductance could be changed over a range of 100:1 whereas a range of 250:1 or more is desirable. The analog could produce two sounds in sequence, and manual resetting of the control system was necessary to synthesize a different pair.

14 5 The synthesizer described in this thesis is a hardware vocal tract analog (hereafter referred to as MINERVA). Vocal tract configurations in MINERVA can be set by computer using an interactive graphics The programs allow long sequences of speech to be synthesized. display. In fact, the maximum length of a speech utterance that can be synthesized is a function only of the computer and the program. The synthesizer operates in real-time*. The synthesizer control programs incorporate any changes that were made in configuration shapes or timing control since the same utterance was last synthesized. The effect of these changes can be heard immediately, as opposed to some synthesizers where a long time elapses before the output is heard. The real-time feature is convenient in designing synthesis rules because a sequence can be heard, changes can be made, and the effect of these changes can be heard immediately. There is always a trade off between flexibility and simplicity of calculations. The software simulated analogs are very flexible in that changes only involve a change in the program, but they require considerable computer time for calculations; so much in fact, that they cannot operate in real-time. Even hardware synthesizers controlled directly by a computer would keep a small computer fully occupied if signals had to be generated every five or ten msec. In MINERVA, a digital hardware interpolator was designed to We define real-time synthesis as follows: immediately upon command, an acoustic utterance from a selected list can be obtained from a sequence of stored control signals, or from stored rules for generating such control signals. Thus synthesis by lowpass filtering of stored samples of a previously sampled acoustic waveform would not be considered real-time synthesis. Although careful examination reveals that our definition is not as precise as one might wish, it seems reasonable and is satisfactory for our purposes.

15 operate between the computer and the vocal tract analog. The computer loads the interpolator with parameter values and rates of change. The interpolator then controls the analog for up to 500 msec, thereby freeing the computer for other tasks. This free computer time is presently used to run interactive software so that the user can easily make on-line modifications in control data. Also, the free time is used to display the spectral characteristics of the synthesized speech. This hardware interpolating scheme vastly reduces the number of times that the analog has to be serviced without detracting from the quality of the synthesized speech. Any small loss in total system flexibility more than compensated for by the additional computer time and memory available for other tasks. The scheme also allows real-time synthesis. This thesis was mainly concerned with development of a good real-time synthesizer and some word synthesis rules. Higher level sentence rules which would be needed for a synthesis by rule scheme were not considered. The purpose of the research in this thesis was to 1. Prove that a vocal tract analog could be built from digitally controlled inductors and capacitors and to develop inductors, capacitors and computer interface circuits for the above analog. 2. Show that the above analog could produce continuous speech in real-time using a small digital computer for control purposes. 3. Study the production of synthetic stops in the initial and middle positions. Production of vowels, liquids and fricatives was also studied. 4. Initiate development of rules needed to control the synthesizer in synthesizing English words. 5. Provide a reliable vocal tract analog for speech synthesis and analysis research. Chapter of the thesis describes the electronic simulation of the vocal tract and briefly describes the various kinds of sounds

16 7 encountered in English. Chapter 3 describes the inductors and capacitors used in the simulation, the hardware interpolator; the excitation sources used, the interactive software and the overall facility. Chapter 4 shows how the vocal tract analog's capabilities were verified in the static mode and also in the synthesis of fricative-vowel and vowel-fricative combinations. Chapter 5 presents a study of stop consonants. Chapter 6 describes the rules used to synthesize 50 English words. Chapter 7 includes a discussion of the speed and memory capacity of a computer needed to control a final version of MINERVA, and some suggestions for further work.

17 8 2. ELECTRONIC SIMULATION OF THE HUMAN ARTICULATORY PROCESS 2.1 DESCRIPTION OF THE VOCAL MECHANISM The major parts of the vocal apparatus are the lungs, trachea, vocal cords, pharnynx, larynx, velum, tongue, oral cavity, nasal cavity and teeth, as shown in Fig. 2.1 a. The vocal tract is an acoustical tube which is non-uniform in area, terminated at one end by the lips and at the other by the vocal cords which separate the trachea from the larynx. In an average adult male the vocal tract is about 17 cm long. The cross-sectional area is determined by placement of the jaw, lips, velum, tongue and pharyngeal wall and can vary from a full 2 aperture of twenty cm to zero for complete closure. The nasal tract is an additional path for sound transmission beginning at the velum and terminating at the nostrils. In an adult male the nasal tract is about 12 cm. in length and has a volume of 3 about 60 cm and the shape does not vary with time. There are three ways of exciting the vocal tract for the production of sound. The most common excitation is voicing. Voiced sounds normally are produced by using air from the lungs and forcing it through an opening between the vocal folds called the glottis. As the air is forced through, if the vocal folds are put in proper position and under adequate muscular tension, they vibrate. This vibration periodically interrupts the air flow, causing roughly triangular pulses to excite the vocal tract. A second way of exciting the tract is by friction. To produce unvoiced friction the glottis is opened so that the air passes freely through it. At some point in the tract there will be a narrow

18 9 Nasal Cavity Velum Lips Teeth Vocal Cords Trachea Oral Cavity Tongue Pharynx Cavity Larynx Esophagus Lungs Chest Cavity Diaphragm (a) Nose Output Mouth Output ^_ Nasal Cavity Velum Vocal / Cords y ~ / \ L. y r Tongue Pharynx Trachea Oral Cavity Hump Cavity (b) Lung Volume r p p/aphragm ^ Force I Figure 2.1 (a) Midsagittal section of the human vocal mechanism (b) Functional diagram of the human vocal mechanism-

19 constriction. If this constriction is sufficiently small a turbulence is 10 created causing a sound similar to white noise. Friction can be Voiced, that is the air passing through- the glottis can be interrupted periodically as in voiced sounds and the turbulence is caused further up the tract. A third method of exciting the tract is to completely close off the tract and build up pressure behind the closure. Opening the closure abruptly, causes a burst of air to come out. A further bufst of air can come from the glottis, when this occurs, the sound is called aspirated. By controlled movement of the vocal tract and proper excitation of the sources, speech is produced. 2.2 HOW SOUNDS ARE PRODUCED The various classes of sounds and how they are produced is described by Flanagan and a short summary is presented here. Sounds which have voicing and no other excitation are the vowels, liguids, glides and dipthongs. The vowels are produced with Voicing excitation and by putting the vocal tract into a configuration which is quasi-stationary for the duration of the vowel. Different vowels are produced by using different configurations of the vocal tract. Typical shapes of the vocal tract for various sounds are shown in Fig The semi-vowels /%/ and /r/ are similar to vowels except the vocal tract is more constricted and the tongue tip is not down. Glides /w/ and /y/ also are produced with voicing but the vocal tract moves during the production of these sounds. The vocal tract is put into a configuration and moves toward the configuration of the vowel which- follows the glide. Dipthongs are voiced sounds that rely on vocal tract motion for their production. The vocal tract changes from one vowel position to another and is in constant motion.

20 t (FOR) 8 (THIN) 3 (1E I i(eve) I (IT) e (KATE) C(MET) S(SHE) h (HE) V (VOTE) se(at) a ( FATHER) 0(ALL) 0 (OOEY) 5(THEN) z (zoo) 3 (AZU*E) U (FOOT) U(BOOT) A (UP) 3 (BIRO) Schematic vocal tract profiles for the production of English vovreb (adapted from TOTTES, Kerr and Guu) Vocal tract profiles lor toe fricative cccsonants of English. The short pain of lines dravn on tbe throat represent vocal cord operation (adapted from POTTXR, Xorr and Caxxir) P(PAT) t(to) k(rtev) Vocal prof2a for tha oasal coasocaata (ju'tar Pvuaa. Xorr and Cull) Fig. 2.2 Vocal tract profiles for the principal phonemes in the English language

21 12 The nasal consonants are voiced sounds. In producing nasals the velum is lowered and the oral tract is usually completely closed off at some point ahead of the velum. The closed oral cavity acts as a side resonator and can influence the sound radiated at the nostrils. Another class of sounds are the fricative consonants there are two types, voiced and unvoiced. The unvoiced fricatives are produced by friction as described in the previous section. The fricative If I is produced with the upper teeth close to the inner surface of the lower lip. The air stream passes through this narrow constriction to cause turbulence. Another fricative, /9/, is produced with the tip of the tongue close to or touching the upper incisors. The'third fricative, /s/, is produced with the tongue raised to approach the back of the gums of the upper teeth. The final fricative, /// is produced with the tongue raised to the front of the hard palate. /h/ is considered to be an unvoiced fricative but is the subject of some controversy. It is thought by many that the turbulence is caused in the larynx. The voiced fricatives are produced by simultaneous voicing and friction described in the previous section. The voiced fricative /v/ and the unvoiced fricative /f/ have essentially the same configuration. Similarly so do /S/ and /9/, 111 and /s/, and /^/ and ///. There is no voiced counterpart to /h/ in English. Plosive or stop sounds are produced by getting a burst as described in section 2.1. Since the vocal tract is closed off for a short time, Voiceless stops are always preceded by a silent interval: In a voiced stop however, during this silent interval there might be a low frequency hum radiating from the throat, not through the oral or nasal aperture but through the walls of the Vocal tract, jaw and Gheeks. The voice source is turned on at some time during the plosive sound or during the transition to the following vowel. Plosives are usually

22 divided into two groups, unvoiced, /p,t,k/ and voiced, /b,d,g/. In English whether a stop is voiced or unvoiced is usually thought to depend on the time the voicing starts, with the voicing starting sooner in voiced sounds than in unvoiced sounds. Plosives /p/ and /b/ are produced by causing the closure to be at the very front of the vocal tract, the lips. Plosives /t/ and /d/ are produced by closing the tract off at the back of the teeth. The last two plosives /k/ and /g/ are produced by closing the tract off further back with the back of the tongue against the palate, although for these two stops the exact place of closure is strongly influenced by the neighbouring sounds, the vowels in particular. 2.3 AN ELECTRICAL ANALOG OF THE HUMAN VOCAL TRACT The human articulatory system is illustrated in Fig. 2.1(b). In an electrical analog of this system current is analogous to air volume velocity, and Voltage is analogous to sound pressure. If the cross-dimension of the vocal tract is appreciably less than a wavelength and as long as the tube doesn't flare too rapidly, the acoustical system can be approximated by the Webster equation c 9t where A(x) is the cross-sectional area, p sound pressure and c the velocity of sound in air. This equation is analogous to the equation for voltage on a transmission line. uniform The vocal tract can be approximated by a series of short right cylindrical tubes of various areas such as the one shown in 26 Fig. 2.3(a). It was shown by Fant that equivalent circuits can be

23 circumference S (3) cross-sectional area A c i G G «_ C 2 ~ " ^ ~ 2 IT Section L A R 5 I -if 0)O/J 6 = 7 k - ' f e p u g : air density : viscosity coefficient of air : radian frequency : specific heat of air at constant pressure c : speed of sound in air n : adiabafic constant of air A : coefficient of heat conduction (b) Figure 2.3 (a) Right cylinder; length I, cross-sectional area A, circumference S. (b) Electrical analogs for a short right cylinder found, for such a tube. Two equivalent circuits appear in Fig. 2.3(b). Inductance L, capacitance C, resistance R and conductance G represent respectively, air inertia, air compliance, viscous friction-at the tube wall and heat loss at the tube wall. An arbitrary constant corresponding to impedance scaling can also be included in the equations. These equivalent circuits will be valid if the cross-dimension of the tract,

24 15 L, i 2 ^ k L s, V L? n c 4 ' c 6 ' c 7 -/2 ;3 14 '/5 I, ^7 6 '/2 Li L> = Cr- 1 k? c H, L 2 = H 2.75 nf, C 2 =5.87nF 470 Q, L r = 270mH rir 4Zw7s t II ft fc/77) m t[ (nf) mi Figure 2.4 Schematic of MINERVA. Elements with double arrows are controlled directly from a digital computer. Capacitors with single arrows are controlled directly by hardware averaging numbers used to control adjacent inductors. Elements C^,, and L are time-invariant. Constants p and c are defined in Fig Constant k is arbitrary. The table shows^the simulated length I for each section, and inductance L and capacitance C which result when corresponding control number B = 1. Also shown for each inductor and capacitor are the respective number of bits n and m in each control number B. The output voltage is v Q (t).a noise source could be inserted in series with any of the inductors. and the length' of the cylinder is small with respect to'a wavelength. By cascading the u or T sections in Fig. 2.3(b) a vocal tract analog consisting of short circular tubes of various' areas can be modelled. 2.4 OVERALL DESCRIPTION OF THE VOCAL TRACT ANALOG The vocal tract analog that was constructed is shown in Fig. 2.4.

25 It consists of 14 cascaded TT sections that simulate 14 right cylinderical tubes. Series and shunt conductors as shown in Fig. 2.3(b) are not included as the losses in the inductors and capacitors are ample. The 16 arrangement of the section lengths is similar to that used by Rosen The first two sections representing the laryngal cavity have a fixed length of 1.0 cm and a fixed radius of.85 cm and.91 cm respectively. Sections 3 to 10 have a fixed length of 1.5 cm and have a variable area that can vary over a range of 64:1 with the radius ranging from.28 cm to 2.24 cm. Sections 11 and 12 have fixed lengths of 1.0 cm and.5 cm respectively and can vary over a range of 255:1 with the radius varying from.14 cm to 2.24 cm. Sections 13 and 14 have a variable area and a variable length because the length of the vocal tract in front of the teeth can change. Lr and Rr simulate the radiation load of the mouth on the Vocal tract. Flanagan has shown that this radiation impedance can be simulated by an inductance in parallel with a resistance as follows: R r ^ i 2! - L r - ^ (2.2) In (2.2) k is the impedance scale factor, r is the radius of opening at the mouth and c is the Velocity of sound. Therefore the resistance is constant and the inductance is proportional to the radius. The glottal end of the tract is excited by a constant current generator which generates: 1) triangle shaped pulses for voicing, 2) noise for aspiration or 3) a combination of both. For production of fricatives and stops, noise sources can also be inserted at any point along the tract to simulate turbulence. The inductors and capacitors should permit simulation of areas 2 ranging from complete closure to 20 cm. Since complete closure represents an arbitrarily large inductance it was decided that the range of elements

26 17 for the front part of the tract should be 250:1 or greater giving an area variation of 250:1 or greater. The elements should be frequency independent over the range 200 Hz to 5000 Hz and of sufficiently high Q to give proper formant bandwidths. Setting the impedance level of the tract involves the following factors: 1. the range over which the simulated area is to vary, 2. the minimum and maximum values of variable inductors available, 3. the minimum and maximum values of variable capacitors available. 2 From Fig. 2.3 it follows that LC = (i/c). Therefore LC is constant for any given length of cylinder and independent of the impedance scale factor k. Since the front sections near the teeth and lips can have a smaller aperture than the rear sections behind the tongue, the variable that is common to all sections regardless of how much they 2 vary is the maximum A. The maximum A was chosen at 16 cm. The maximum A sets the minimum L and the maximum C. From the inductors and capacitors that are described in the following chapter, a reasonable value for the maximum C was 57.6 nf. For I = 1.5 cm this gives the minimum L a value of 33.3 H. Using these values, which are realizable with the simulated inductors and capacitors, the table of values given in Fig. 2.3 can be developed. These values give the impedance scale factor k a value of x 10 gm cm sec ohm From equation 2.2 Rr * 460fi Lr = 79.5 mh x r. For R, 470Q was used.

27 TIME-VARYING CONDUCTOR 3. HARDWARE AND SOFTWARE To simulate time-varying inductors as described in the next sections, a digitally controllable time-varying conductor was th needed. The circuit chosen appears in Fig. 3.1(a). If the k switch th is closed when b^ = 1 and the k switch is open when b^ = 0, then; G = gb where n-1, B = E b, 2 k=o and G is the conductance between terminals x-x'. If the switches are opened and closed in proper then any time function such as G(t) in Fig. 3.1(b) can be time sequence, approximated arbitrarily closely by a stepwise function G(t), as the following argument shows. In Fig. 3.1(b), r(t) is a piecewise linear approximation to G(t), and s(t) is a stepwise approximation to r(t), the step height is g. Errors between G(t) and s(t) result from resistor inaccuracies. Let = G - r, ~ r-s > 3 = S- G and e - G-G. Since E = e_ + e~ +, e< e 1 1 +[e + e^l ' ' Let G m and G' be the maximum magnim dg tudes of G(t) and - j ^ respectively for a $ t < b. For a ^ t < b e i(t) * t-a (G(t) - G(a))/(t-a) - (G(b) - G(a))/(b-a) It follows that e J <: 2TG'. Since (2 n 1 l -l)g=g and e U g» lej $ G / 1 1 m m 2 1 ' 2 m (2-1). Let Ag k and AG be the variation in g^ and G about their nominal values. n-1 AG= (GFFI/C2-1)) N z (g k /g) ( A g k /g k )b k k=o if Ag k /g k is small. If Ag k /g k sa for all k, then j E 3 < G T N A» A N D

28 19. ~~I Time-varying g <L2y.y4y---y2 n ~'g j-*- Conductance 'Switch j Digital n-i i Interpolating ' i System Direction-of- Count Control Signal BitO Bit I Bit 2-"-Bit n-1 V. Digital Computer or other Digital Storage Device ( 3 ) L. Clock Puise Generator Figure 3.1 (a) Discrete-state time-varying conductance and its control circuitry, (b) Stepwise approximation G(t) to continuous time function G(t). All steps in s(t) are of equal height g. Any difference between s(t) and G(t) results from inaccuracies in resistors 2^ in (a).

29 20 28 Figure 3.2 Circuit of Riordan for simulating time-invariant inductor. If the element denoted by Z^ is a time-varying conductance, if Z, Z^ and Z,. are resistive, and if is a capacitor then impedance from 1 to ground is a time-varying inductance. e(t)/g I $ 2T G'/G + l/(2 n 1 -l) + A m 1 mm It follows that e(t) can be made arbitrarily small by making T, l/n, and A small. 3.2 TIME-VARYING INDUCTORS As discussed in the previous chapter, time-varying inductors were needed. These inductors had to be variable over a range of 250:1, had to be frequency independent over the range 200 Hz < f < 5 khz and 27 had to have Q > 25. In 1967 Donaldson and Wickwire developed such an 2 8 inductor from a time-invariant inductor proposed by Riordan. The basic circuit is shown in Fig The input impedance between terminal 1 and ground is in Z 1 Z 3 Z 5 Z 2 Z 4 (3.1) If Z,, Z OJ Z. and Z c are resistive and Z_ = -~ then Z. =, which 1' Cs in G ' simulates an inductor where K is a constant.

30 21 Since a floating inductance rather than an inductance to ground was desired, a symmetrical circuit such as the one in Fig. 3.3(a) had to be employed. If the gains of the differential amplifiers are large and the admittance of C is large compared to then v(t) = L(t) ^"^^ W h e r e L(t) = G 4 C 2 /G 5 G 3 G 1 (t) Since v(t) = dl(t)i(t) a true time-varying inductor, the circuit simulates an inductance if L(t) ^^ t^ <<i(t) ^L(t), if q ± s constant dt dt i and one of the pairs G Gl, G,G!or G C G' is varied simultaneously then v(t) = ( ( t t^ dt ). With i(t) = Icos(2TTft) and A and I as defined in Fig. 2.3 i(t)dl(t) / L(t)di,(t) = JL da/jl dt / dt 27rfA dt ' and was small over the frequency range of interest 200 Hz f $ 5 khz unless A/& changed suddenly and by a large amount. Since the assumptions of the analog were questionable during this time and since much more elaborate circuitry was necessary to vary the conductors simultaneously, the circuit of Fig. 3.3 was used. One bit in the time-varying conductor is shown in Fig. 3.3(b). With G^(t) in Fig. 3.3(a) constant, the circuit simulates a linear time-invariant inductor. Let L be the inductance which results when control number B=l. Note that since G^ = gb, L = L/B. With L = 10.OH (C =.01 uf) the inductor's Q exceeds200 for the range of frequencies 100 Hz f 20kHz and 10 B 255 as shown in Fig. 3.4(a). The Q exceeded 50 for 100 Hz f 5 khz and 11$ B < 9. Similar behaviour of Q was observed for L = 1.0 H(C=.luF). For = 10.OH and again for L = 1.0H, LB/L deviated appreciably from unity over the frequency range 200 Hz ^ f ^ 10 khz only for B ^ 75, as shown in Fig. 3.4(b and c). Fortunately, B ^ 63 for most

31 22 10K 4.7K 'v(t) Time-varying Conductor G,(t) \ Digital Inter- Ipohiing System 2N5I63 2H5IS3 M?! MPS3638 IM G8K i t ^Af5/63^- 7? rib; Figure 3.3 (a) Ungrounded time-varying inductor. A grounded inductor results if A and B are grounded, (b) Series combination of the k resistor and switch in G..(t). When v, = -6 volts, b. = 0. 1 k k When v^ = 12 volts, b^ = 1.

32 to 10 6 c 4 o c: 2 sy &* 1 w lit O.b 0.4 / 1 \ i i / 1 ' \' i y> i V 7 \j o <&> ] >zoo! j of- region interest vocal analog for tract <4> -S <OQ o (g) * 9 Bindry Control Number QQ & 1.2 g C B'l, 10 / B*255 f OA d 10 (fo) Frequency f (KHz) B-IO, 100, / SH B=I A B-100 t B* W 255V OA 0.6OBI Frequency f (KHz) (C) ure 3.4 (a) Q of steady state inductor plotted against control number B and frequency f. (b) Steady state inductance for the circuit of Figure 3.3 vs. frequency f and control number B. L equals the inductance for B = 1. L = 10H, (c) Same as (b) for L = 1.0H.. '

33 24 inductors in Fig. 2.4 and B < 75 for the remaining inductors most of the time. The Q of radiation inductor Lr in Fig. 2.4 was between 5 and 10 for 200 Hz f 5 khz independent of B. In this case a low Q can be tolerated as the inductor was in parallel with a 470fi resistor. Ratio IB was within 3% of its theoretical value for 200 Hz $ f $ 5 khz for 1<B<15. L Transient tests were done and are described in Wickwire's 29 thesis and 27 in Donaldson and Wickwire 3.3 TIME-VARYING CAPACITORS If the circuit of Fig. 3.2 is used with resistive and Z^ capacitive then a simulated time-varying capacitor is realized, since i( t ) = dc(t)v(t) w. t h c ( t ) = R G ( t ) C f (3.3.1) at 29 Although this circuit was tested by Wickwire, when it was connected to the simulated inductor the combination was found to be unstable in the sense that the outputs of the operational amplifiers oscillated between saturation in the positive direction and saturation in the negative direction. A more stable capacitor was then realized by modifying the 30 circuit in Fig. 3.5 which was originally used by Antoniou. The modification resulted when the time-varying resistor R-j(t) in Fig. 3.6 replaced the fixed resistor Z^ in Fig In Fig. 3.6, i(t) and V(t) are related by as is shown below: V(t) R^R, V (t) = - 02 U ; v «(t) V(t)-V (t) R v o i ( t ) = 0 2 OF) V w R 0 (t). c 5 d(v(t)-v 01(t)) i ( t ) _

34 25 Figure 3.5 Circuit used to simulate capacitor. Time-varying capacitor is realized if is a time-varying conductor. if then i(t) R 3 = N(OG Kdv(t)C(t) dt a n d N ( t ) C5 " C ( t ) Fig. 3.7 shows the measured capacitance C and quality factor Q vs R^. Fig. 3.8 shows C and Q vs control number, B, when the resistor ladder with FET switches was substituted for R^. In Fig. 3.8 good linearity of C with respect to B is observed. The Q was highest for mid-control numbers and for higher values of C^ and was lowest for small numbers and small C^. The Q was always above 25 and was usually greater than 100. Eig. 3.9 shows Q plotted against frequency for various values of R^ and C^. The input.capacitance (not shown) was within 2% of the calculated value for 200 Hz S f S 10 khz. The frequency range of interest is between the two arrows in Fig The Q is always above 30 except when C f = 2.2 nf, R 3 = 100 kft and f $ 500 Hz. Because only one capacitor of the 12 variable capacitors in the vocal tract analog has C^ = 2.2 nf the circuit in Fig. 3.6 was considered acceptable for

35

36 Variable Resistance R q (Ohms) Figure 3.7 Input capacitance C and quality factor Q vs. variable resistance R for capacitor in Fig. 3.6 (1) C f = 2.2nF, (2) C f =4.7nF, (3) C f = lonf, (4) C f = 22nF. The measurements were taken at a frequency of 1 khz. use in MINERVA. 3.4 DIGITAL INTERPOLATOR It was desirable to minimize the time required for the computer to calculate and transfer control numbers to the time-varying resistors in MINERVA. Accordingly a peripheral hardware digital interpolation system was designed with the result that the computer needed to supply control numbers only at the break points, For each inductor or capacitor to be controlled a register of semi-conductor memory, a digital comparator, anup-down counter and a variable rate clock were needed. Fig. 3.1 shows the control scheme used to control each element in a

37 28 Binary Control Number B Figure 3.8 Input capacitance C and quality factor 0 vs. binary control number B for capacitor in Fig (1) C^=2.2nF s (2) C^= 4.7nF, (3) C f = lonf. The measurements were taken at a frequency of 1 khz. digital piece-wise linear fashion. The value of the element was controlled by an up-down counter, which counted at a rate determined by the variable rate clock. The output of the up-down counter was connected to a digital comparator which compared the number in its memory with the number in the up-down counter. If the number in the up-down counter was larger than the number in the memory the comparator enabled the do\m count and the counter decremented. Similarly if the. number in the counter was less than the number in the memory, then the counter incremented. If the numbers were equal then the counter was disabled. The variable rate clock consisted of a down counter and a memory register. The memory register contents were loaded into the counter which

38 29 ion* JOOHZ m U z iokhz: Frequency Figure 3.9 Quality factor Q vs. frequency for capacitor in Fig. 3.6 for various values of C f and R,. (1) C f =2.2nF, (3) C f =10nF, (6) C f = 22nF. For C(6) and = lokfi Q>1000 and is not shown on the graph. counted at a fixed rate. When the counter reached zero, an output pulse was generated and the contents of the memory register were loaded into the counter again. If the number C was loaded into the memory register and the clock was driven with a square wave of period T, the output would be a pulse train of period Cx. This.period would be maintained until the memory register in the clock was changed by the computer. This control scheme utilized the idea of - target configurations for the vocal tract,. The vocal tract would move from one position to another and then hold that position. For example, assume that the number a was in the up-down counter and the number b was loaded into the digital comparator's memory register at time t^, and at that same time the number T/( b-a ) was loaded into the clock. The up-down counter and hence the element would reach the

39 7* Figure 3.10 Control number N as a function of time using up-down counter, comparator and variable rate clock. value of number b at time th-t, and would hold that number until the memory register in the comparator was changed at time t, as shown in Fig The basic period for each variable rate clock was.125 msec, therefore the maximum rate of change was one step every.125 msecs. The slowest rate was one step every 128 msecs. An interrupt clock was built by modifying a variable rate clock so that the counting could be disabled and enabled. The enabling pulse to the interrupt clock synchronized the astable multivibrator. When the counter reached zero an output pulse was communicated to the digital computer. The output pulse also disabled.the astable multivibrator. The effect of the interrupt clock was as follows: if the number C was loaded into it, the computer would.receive an interrupt Cx later where x is the basic period of the astable multivibrator. In MINERVA x was.5 msecs and C was a 10 bit number, therefore times of.5 msecs to 512 msecs could be set. The purpose of the interrupt clock was to allow the computer to calculate the time of the next interrupt and to load the interrupt clock. When MINERVA required the next service, it would request it by raising an interrupt.

40 GLOTTAL SOURCE As noted in Chapter 2 vocal cord excitation is caused by periodic interruption of the flow of air through the trachea. The source has a high acoustical impedance and its electrical analog is a current source. In MINERVA the glottal source could generate a periodic buzz to simulate voicing and/or an aperiodic noise to simulate aspiration. A control word allowed voicing only, aspiration only or both simultaneously. The circuitry for the buzz source which produced a periodic triangular shape pulse is shown in Fig. 3.11(a). If G(t) was a digital time-varying conductor as described in section 3.1, then a voltage v(t) as shown in Fig. 3.11(b) appears at the output if G(t) increases with time. If square pulses p(t), as shown, excite the base of transistor Q then a train of pulses of increasing amplitude would appear at the collector of transistor Q. The pulse generator was either a fixed frequency pulse generator or a variable rate pulse generator which will be described later. The fixed frequency pulse generator is an astable multivibrator. The pulse shaping filter is a lossy integrator which will give the waveform i g (t) shown in Fig. 3.11(b). This Xirave is then summed with the aspiration and applied to a voltage to current converter, making the glottal source appear as a current source. The spectrum of the buzz waveform is shown in Fig along 15 with the spectrum used by Stevens et al in 1953 on a static vocal tract analog. The two spectra are within 4 db when 100 Hz f $ 5 khz. Stevens et al chose their spectrum by listening to synthesized vowels and adjusting the-source spectrum for optimum vowel quality. To simulate changing pitch during speech production a variable period pulse generator was added to the glottal source and fixed or variable

41 V A/ariable Gain Amplifier 1 j -Pulse Shaping Fitter z Summing Amplifier ii T ^ M V SI Pulse [ Generator p(t) _ -J Voltage-to - current - Converter Vocal Trsct Analog T > + raj f s COi > 7//77e Cons/ant t-rc \ 1 /\ V Figure 3.11 (a) Glottal excitation source. (b) Waveforms. In MINERVA a = 200ysec, g.= 8 msec, W =0.5 msec. Pulse generator p(t) is periodic, with period ft. to N5

42 Figure 3.12 Spectrum of voicing source used and spectrum of source used by Stevens et pitch was selectable by a switch. The pulse generator, shown in a block diagram in Fig. 3.13, was a variable rate clock, described in Section 3.4, in which the memory register has been removed and replaced by the output of an up-down counter. After each output pulse the modified clock's counter was loaded with the contents of the up-down counter. If the period of the output of the variable rate clock toggling the up-down counter is long compared to the period of the astable multivibrator in the modified clock, then the effect will be pulses coming at a slowly varying period as shown in Fig Again a comparator, clock, and up-down counter combination were used to set the new period of the pulse generator and the rate at which to move to the new period without further service from the computer. Although the pulse generator had an effective frequency varying from 33 Hz to 8 khz the range 100 Hz to 150 Hz was used during most speech synthesis studies.

43 Figure 3.13 Variable rate pulse generator used to simulate changing inflection at voicing source.

44 J J n n n n fl Variable Rate n Clock Pulses Figure 3.14 Variable rate clock pulse produced by circuit in Fig NOISE SOURCE Pseudonoise was used to simulate aspiration and friction. 31,32 The pseud.onoise generator, suggested by A. C. Davies ', and shown in Fig. 3.15, generated the noise by loading a 6 bit digital to analog (d-a) converter with the binary output of a 15 bit m-sequence shift register which had a 1 MHz clock rate. The noise had a uniform amplitude probability density function whose short term spectrum was effectively white over the range 0 < f < 10 khz. The output of the d-a converter was connected to a variable gain amplifier, like the one in Fig. 3.11, whose gain was varied by varying G(t). The gain G(t) was a trapezoidal function except when voicing and friction occured together. In the latter case G(t) varied at the voicing frequency between bn(t) and the larger of {0:bn-bs} as shown in Fig The numbers bn(t) and bs(t) were 6 bit binary numbers proportional to the ampltidue of the noise source and the voicing part of the glottal Source respectively. This alternation simulated the periodicity of the air in the tract coming through the vocal folds and passing through a constriction which caused turbulence. In MINERVA the location selector was used to select the proper transformer for inserting the noise into the vocal tract.