Digital-Formant Synthesizer for Speech-Synthesis Studies

9.4, 9.8 Received 1 November 1967
Diital-Formant Synthesizer for Speech-Synthesis Studies
LAWRENCE R. RABINER
Bell Telephone Laboratories, Incorporated, 3fltray tlill, A;ezc, Jersey L7974
Recent work on speech synthesis by rule has led to the design and evaluation of a new digital formant
synthesizer. The synthesizer was simulated on a digital computer at a 20-kHz sampling frequency. The
improvements over previous synthesizers include: an excitation sonrce for the unvoiced compmeut of
voiced fricatives; a method for producing a voice bar for the closure period of voiced stops; and a simple
desigu for the higher pole correction network that exploits the properties of sampled data systems. Spectrographic
examples of synthetic speech are shmvn in order to illustrate these improvementS. A brief discussion
of some advantages and disadvantages of serial and parallel synthesizers is presented.
INTRODUCTION
N recent lears, an extensive amount of time and ability of a terminal analog synthesizer to produce high-
effort hats .gone into designing and evaluating qnalily natural-sounding speech has been demonstral ed
machines capable of converting tow-information-rate repealedly, u,-ø thus motivating efforts to increase the
control signals into speech. -4 hnprovements in channel capability of this machine and to make it more versatile.
and formant vocoders have been in part responsible for Its potential as a speech-research tool is high.
this resnrgence of interest in speech synthesizers. :,-7 This paper is concerned with at digital simnlation of
Furthermore, tire general availability of both large and a serial terminal-analog synthesizer. The synthesizer
small digital computers 8. and advances in digital and was sinrelated on a large computer (GE 45) usin
analog hardware have contributed in large meastu'e to FORTRAN I¾. The general structm'e of the synthesizer
the widespread interest in speech synthesizers. in relation to the production of various classe
Among the more common speech synthesizers is the sounds is examined closely, and improvements over
terminal-analog synthesizer. Besides serving as the previous synthesizers are discussed in detail. These imsynthesizer
for a formant vocoder, a terminal-analog provements include a voiced fricative excitation netsynthesizer
is often used to generate material for experi work, a new design for the higher pole-correction network,
md a simple method of producing a voice bar for
R. S. Tomlinson, "SPASS An Improved Terminal-Analog voiced stop consonants. A discnssion of some ad-
Speech Synthesizer," J. Acoust. Soc. Am. 38, 940(A) (1965).
vantages and disadvantages of parallel and serial
G. Fant, J. Martony, U. Rengman, and A. Risherg, "OVE II
Synthesis Strategy," Paper I:5, Speech Conminn. Seminar, realizations of a terminal-analog synthesizer conclude
Stockhohn (1962).
this paper.
a j. L. Flanagan, "Note ou the Design nf Terminal-Analog
Speech Synthesizers," J. Acoust. Soc. Am. 29, 306--310 (1957).
* C. H. Coker and P. Cummiskey, "On-Line Computer Control
of a Formant S.xmthesizcr," J. Acoust. Soc. Am. 38, 940(A)
(1965).
s C. H. Coker, "ReabTime Formant Vocoder, Using a Filter
Bank, a General-Purpose Digital Computer and an Analog Synthesizer,"
J. Acoust, Soc. Am. 38, 940(A) (1965).
B. Gold, "Techniques for Speech Bandwidth Compression,
Using Combinations of Channel Voc(>ders and Formant Vocoders,"
J. Acoust. Soc. Am. 38, 2 10 (1965).
M. R. Schroeder, "Vocoders: Analysis and Synthesis of
Speech." Proc. IEEE 54, 720-734 (1966).
8 p. Denes, "On Line Computing in Speech Research," Proc.
Intern. Congr. Acoust. 5th Liege (1965), Paper A23.
822 Volume 43 Number 4 1968
ments in speech synthesis and perception.0.. The
I. GENERAL DESCRIPTION
of speech
A termired-analog synthesizer models the speech producing
mechanism, including the vocal tract, excitation
0 K. Xakata, "Synthesis of Nasal Consonants by a Terminal
Analog Synthesizer," J. Radio Res. Labs. 6. 243 254 (1959).
o L. Rabiner, "Speech Svnthesis hv Rtde l An Acoustic Domain
Approach," Bell System 'lrech. J. 47,' 17-37 (1968).
u j. N. Holmes "Notes on Synthesis \V0rk," Speech Transmss
on Lab. Qnart. Progr. Rept.', Stockholm (April 1961).
2 W. J. Strong, "Machine-Aided Formant Determination for
Speech Ssmthesis," J. Acoust. Soc. Am. 41, 1434 1442 (1967j.

DIGI'FAL-FORMANT NYNTHE.IZER
f T t-
I"m. 1. Blck (liegram .f sl)cech synthesizer.
sources ,red r,tdiation impedance. The fieorv of repre- tontrol sigmd is set to t,ttc the pitch impul_-e generator
senLing vocal-tract lelmvior in tel-mS O[ ils Irtnsmis- Otltpnl to the shaping nehvork.
sion properlies, I) a trausfcr fnnclion has been

L. R. RABIXER
F o COMPOiEI T Figure 2 shows the relevant details of this network.
OF VOICED
(Certain components of the synthesizer of Fig. 1 have
FILTER NET .ORK
been omitted from Fig. 2 for simplicity.)
The unvoiced excitation is produced as follows. The
pitch pulses excite a resonator tuned to the first forerant
of the voiced cmnponent of the fricative. This resonator
Bt RESONATOR
is the first-order pproximation to the transfer function
of voltame velocity (the signal of interest in Fig. 2) from
the glottis through the point of constriction of the vocal
tract. A threshold level (Vn) is subtracted from the
output of the resomUor, and the restlit is half-wave
rectified. These operations model the physical fact that
turbulence is not produced until the volume velocity of
the airflow exceeds a threshold value.
'['he output of the half-wave rectifier modulates the
GENRATCR NOISE J b'VOlCEO outpnt of the noise generator, producing a pitch-
COMPONEN1
OF OlCES synchronous excitation for the unvoiced component of
the fricative. The final unvoiced component is produced
l'lc.. 2. Excitation netxmrk for voiced fricatives. by feeding this excitation into the fricative network
(i.e., the lower branch of the synthesizer). The voiced
generator, a zain control, a multiplier, a frication pole component is prodneed by exciting the fore,ant network
and zero network, and a shaping network. The gain by the pitch pulses.
control AFRIC provides means for controlling the over- Figure 3 shows spectrograms of synthetic and natural
all unvoiced output level.
versions of the voiced fricatives/z/and/5/- The syn-
The input to the multiplier from the voiced-fricative thetic speech was produced eutirelv bv rule. m The
excitation network is constant except during a voiced natural speech is presented for comparison purposes. A
fricative. For a voiced fricative, the input to lhe multi- careful examination of these spectrograms shows the
plier is a signal that pitch s3nchronously modulates effects of the pitch-synchronous modulation during the
the unvoiced input. The details of this technique are fricatives fro' bolh the synthetic and the natural speech.
explained in the next Section.
The presence of the low-frequency voiced components
The fricative pole and zero network is simply a cas- for both the synthetic and natural versions can also be
cade of a complex-conjugate pole-pair resonator and a
seen in Fig. 3. The major difference between the comcomplex-conjugate
zero-pair resonator. The pole and
zero positions are variable (FPOL, FZER). The shaping
panion spectrograms is that the lower formants of the
network i, used to provide high- and low-frequency natnral speech are weaker than the lower formants of
emphasis, a, well a,to acconnt for radiation characteristics.
A cascade of a differentiaLor and a low-frequency
resonator is used here.
The middle branch of the synthesizer is used to produce
a voicebar for the closure interval of voiced stop
consonants. This branch consists simply of a gain
control (AVB) that adjusts the level of the pitch pulses
6- {
produced at the output of the ,aptrig network in the
upper branch.
The output ,eech is produced by adding the output
from each of the three branches.
II. VOICED FRICATIVE EXCITATION NETWORK _ I
The voiced-fricative excitation network that connects s-
the output of the pnlse-shaping filter to the lower
branch of the sx nthesizer is used to model the produc- .z a-
I,,
tion of the unvoiced component of voiced fricatives) :-'
HUMAN ,%øEEC H
O. Fujimura (personal communication, 1067) has indicated
o
that G. Rosen used a modulation scheme to produce voiced frlca- o-
tires on D \VO [G. Rosen, "Dynamic Analog Speech Synthe- aS u
sizer." J. Acouq. Soc. Am. 30, 201-210 (1958)]; however, no
x titten record of his techniques appears to be available.
I'h(. 3. Spectrographic examples of voiced fricatives.
824 Volume 43 Number 4 1968
P, HZ 3-
2-
HUMAN SPEECH
! I
SYNTHETIC & Z ae -- SYNTHETIC

DIGITAL FORMANT SYNTH ESIZEI(
the synthetic versions, indicating a need for low-fi-equency
de-emphasis.
The intelligibility of the voSced fricatives was measured
in a formal test, ils part of a larger-scale evaluation
of synthetic consonants./" The relevant resttits are that
/z/ and /5/ were identified carteeth' 100% of the time
in prestressed and poststressed position in VCV tests.
There were 15 possible responses for each stimnlus in
these tests. The voiced fricatives /v/ and .'8./ are not
included in this discussion because the unvoiced component
was not found necessary for their synthesis.
Hence, they were synthesized in the same way ils vocalic
sounds.
No formal tests were conducted to determine the
quality of the synthetic voiced fricativcs, bul informal
comments by listeners were highly favorable.
III. HIGHER-POLE CORRECTION NETWORK
I"m. 4. Pole positions for a five-pole 10-kHz synthesizer.
+24
- t 2 ST RESPO 'NSE 0 ' AC
0
I 2 TUBE 5 POL ANA
36
48
O I 2 3 4
FREQUENCY ( k H Z ) --'
l"JO. 3. Digital and aliatog approximations to transfer function
of an acoustic tube, open at me end and closed at the other end.
sampling fi'equency i.e., the pole at 500 Hz represents
The higher-pole correction network is used to com- analog poles at 500, 10 500, 20 500 Hz, etc. Hence, the
pensate for the effects, Oil the low-frequency- spectrum, infinity of higher-order poles is guaranteed by using this
of the missing higher-order poles in the usual approxi- method. The density of the higher-order poles is also
mation to the vocal-tract transfer function. In an correct, thus automatically ensm'ing the necessary
analog synthesizer, a network approximating the trans- higher-pole correction. The efi'ectiveness of the digitill
fer function of the higher-order poles is used. In a digital higher-pole correction is shown in Fig. 5. Here we see
synthesizer, however, the design of the higher-pole car five-pole digital and five-pole anah)g approximations to
rection is relativeh' simple. Additional resonators, or the transfer function of a straight acoustic tube modelformant
networks, placed at appropriately chosen fre- ing the vowel/'a/'. No tnalog higher-pole correction is
quencies, provide the necessary higher-pole correction. used other than the additional fixed formants at 3500
To illnstnte this method, consider it synthesizer and 4500 Hz. '['he bandwidths of the digital ired analog
with a 10-kHz sampliug frequency. Assume the vowel poles are increasing with frequency- tmodeling the ob-
/a/is to be synthesized and the pole positions are 500, served data for speech), whereas the bandwidths of the
1500, 2500, 3500, 4500, ... (2n-+-l) $00 Hz. As shown poles of the tube transfer function are constant. The
in Fig. 4, we place the three resonators of the formant high-frequency behavior of the analog approximation is
network at 500, 1500, and 2500 Hz. The higher-pole significantly different from the tube trinsfer function,
correction network consists of resonators at 3500 and indicating the need for additionM higher-pole correction.
4500 Hz. Owing to the periodicity, in frequency, of The advantages of the automatic higher-pole correction
sampled data systems, the pole at any' frequency repre of the digital approximation are apparent.
sents an infinity of poles with it spacing e(lual to the The modifications of the higher-pole correction network
for a 20-kHz sampling frequency are straightforward.
Instead of inserting fixed resonators at just
z - PLANE
35(10 and 4500 Hz, additional resonators at 5500, 6500,
7500, 8500, and 9500 Hz are necessary. There will be 10
poles in the upper half-plane of Fig. 4, instead of five as
for it 10-kHz syuthesizer; however, the poles are at the
same frequencies as before. For a 20-kHz synthesizer,
TAB1,E I. Resonator frequencies and bandwidths.
Resonator C[: (ttz) Q BW (Hz)
1:4 35011 2(I 175
I":, 4500 16 281
b' 5500 12 458
I"7 6500 9 722
l"a 7500 6 1250
Fo 8500 4 2125
P'o 9500 2 4750
The Journal of the Acoustical Society of America 825

kHz
kHz
4--
2--
5--
4--
3--
2--
1--
O-- !
El U T
T
1
I
L. R. RABINER
t t'1 I
, I
A N S I S T O R R A D IO
S E L DOM GET
I'm. 6.1..SI)ectrograms of synthetic speech.
R I C H
the b,ndwidfi-, of the additional poles are.reater than citin the upper branch of the sx nthesizer !)v the pitch
the bandwidth of the similar poles for a 10-kHz sxn- impulses
thesizer. The bandwidths used are seen iu Table L The Voiceless stop consonants are produced using the
values chosen were determined from data by l)unn. s upper and h)wer branches of the synthesizer. The time
interval following the period of closure can be broken
IV. SYNTHESIS CHARACTERISTICS
SPEECH SOUNDS
OF THE down into three senents. During the firsl seg-ment,
there is a short burst of friczttion uoise. This uoise is produced
using the lower branch of the synthesizer. The
In order to produce the various clases of speech
output of the fricalion enerator is multiplied by m exsounds,
only certain portions of the synthesizer need
ternal ,unplitude control (AI"RIC) and then multiplied
generally be used at one time. Vowels, semivowels, by a fixed constant. The output of the second multiliquids,
and glide are produced using only the upper plier is fed into the fricative pole and zero network, and
branch o[ the x nthesizer. The pitch-impulse generator
then into the shaping network.
is the source of excitation for these sounds.
The second segment of a voiceless stop is the aspira-
Voiced stop consonants are produced using both the
tion ptmsc. Aspiration i. produced bv gating the output
upper and middle branches o[ the synthesizer. The voice of the frication generator to the upper branch of the
bar is produced bx gating the output of the pitch- synthesizer. The third segment of the voiceless stop,
impulse generator to the shaping network ,red lhen the voiced transitional phase, is px oduced in exactly the
multiplying the output of thi.,, network by an extern,'fi same m,mner as for voiced stops.
amplitude control (AVB). This provides a low-fre-
The voiceless fricatives are produced in the same way
quency 1ow-enerzy waveform whose characteristics are as Ihe [rication noise for voiceless stop consonanls. As
similar to voice bars observed in real speech. The
shown above, a voiced fricative is synthesized as two
tnmsitional sebment of voiced stops is produced by exseparate
components. The voiced cmponent is produced
in a manner similar to the voiced components of
olher speech sounds--i.e., by excitin the upper branch
]I. K. Dunn. "Methods of Xleauring Vowel Formalit Bandwidths,"
J. -cout. ,qoc. Am. 33. 1737 1746 (1961).
826 Volume 43 Number 4 1968

DIG[TAL-FORMANT SYNT[fES[ZER
of the synthesizer by pitch impulses. The unvoiced component
is produced by modulating the frication generator
output by a pitch-synchronous waveform--the
output of the voiced-fricative excitation network. The
modulated waveform is then passed through the lower
branch of the synthesizer and added to the voiced component
to produce the voiced fricatlve. An/h/ sound
and whispered speech are produced in a manner similar
to that for aspiration.
A. Spectrographic Examples
In order to illustrate the capabilities of the synthesizer,
we show, in Fig. 6, spectrograms of two synthetic
utterances. The control signals used to drive the synthesizer
were determined by rule using an auxiliary progrant.
Therefore, there are no original speech samples to
present for comparison purposes.
I I
These spectrograms exhibit the various classes of
speech sounds. The upper spectrogram shows both
voiced sounds and unvoiced sounds. Of most interest in
Fio. 7. Parallel terminal-analog synthesizer.
this spectrogram are the phoneraes in the word lransis- ticated parallel synthesizer would account for formant
lor. I)uring the/t/, both the frication noise burst and phaseø-0).
aspiration are visible. The effects of the nasal pole and The serial synthesizer presents several advantages
zero for/n/' are noticeable in the region of 2 kHz, where over the synthesizer of Fig. 7. Individual amplitudes of
there is little energy present. Finally, a comparison is each of the resonances do not have to be determined for
available between the voiced fricative /z/ and its voice-
, /
a serial synthesizer. For research on speech synthesis by
less counterpart /s/. The effects of the pitch-synchronous
modulation are easily identified by the vertical
rule, this is a significant advantage, as it reduces the
striations through the noise of/z/. Furthermore, the complexity of the rules. The vowel spectra from the
lack of a low-fi'equency voiced spectrum for/s/is easily parallel synthesizer of Fig. 7 contain extraneous zeros,
observed.
whereas a serial synthesizer produces spectra containing
The lower spectrogram exhibits two interesting features.
The voice bars of /d/ in "seldom" and /g/ in
only poles. These zeros generally fall at frequencies
between the resonances and may be perceptible, and
"get" appear as low-frequency vertical striations during hence a corrupting factor. Because of the presence of
their respective stop gaps. The voice bars appear as zeros in the spectra, parallel synthesizers generally have
light lines indicating that they are of low energy. no higher-pole correction networks, relying instead on
Finally, the affricate /ti/ of "rich" appears as a stop the zeros to provide the low-frequency emphasis. This
gap followed by a burst and ending in a period of frica- is an nnreliable technique because: the positions of the
tion energy. The frication energy is similar to the energy zeros move as the resonances are varied.
of the /'/ phoneme. An affricate is synthesized as a
The parallel synthesizer has two advantages over a
stoplike sound followed by a fricativelike sound on this
serial synthesizer. Noise generated in the parallel synsynthesizer.
thesizer propagates additively rather than multiplicatively,
as in a serial synthesizer. For simulations where
B. Parallel versus Serial Realization
small register lengths are required, this advantage is
particularly valuable. For a given signal-to-noise ratio,
The discussion in this paper has been concerned pri- signal sizes in a parallel synthesizer are smaller than in a
marily with a serial tenuinal-analog synthesizer. A serial synthesizer. The second advantage is the ability
parallel synthesizer is generally realized by gating either
of a parallel synthesizer to reproduce consonant spectra
of two sources (a noise or pitch-impulse generator) to a
accurately through independent control of formant
series of parallel branches, each containing a resonator
amplitudes. Hence, for synthesis from spectrograms or
and a gain control, and adding the output of each
in a formant vocoder, this degree of freedom can be of
branch to produce the speech. Figure 7 shows a simple
realization of a parallel synthesizer l,ø (a more sophisgreat
value.
J-. Holmes, I. Mattingly, and J. Shearme, "Speech Synthesis
by Rule," Language and Speech 7, 127-143 (1964).
PITCH
GENERATOR [
EECH
J. L. Flanagan, "Recent Studies in Speech Research at Bell
Telephone Laboratories (II)," Proc Intern. Congr. Acoust., 5th
IJage (1965), Paper A22.
The Journal of the Acoustical Society of America 827

L. R. RABINER
The choice of a parallel or serial synthesizer is determined
primarily by its application. For research in
speech synthesis by rule, the serial synthesizer has been
very useful.
have been included to demonstrate the capabilities of
the synthesizer.
ACKNOWLEDGMENTS
V. SUMMARY
I am indebted to Dr. Cecil Coker of ]3ell Telephone
Laboratories and Professor Osamu Fujimura of the
A new design for a digital terminal-analog synthesizer Research Institute of Logopedics and Phoniatrics for
has been presented and discussed. The new features many valuable discussions on the design of a speech
include a voiced fricative excitation network, a method synthesizer; to Dr. _[ames Flanagan, Professor Kenneth
of producing a voice bar, and a simple higher-pole cor- Stevens, and Professor William Henke for sugõestions
rection network based on the properties of sampled data and ideas about speech synthesis; and to Dr. Harry
systems. Spectrographic examples of synthetic speech Levitt for valuable criticisms of the manuscript.
828 Volume 43 Number 4 1968