AUDIO
OUT
AUDIO OUT
L
R
By Jake Rothman
Analogue Vocoder – Part 1
T
ime for a new music project!
Many of you are probably familiar
with the vocal effect of an electronic
instrument called the ‘vocoder’. It’s a
fascinating piece of electronics, and over
the next few months we will design and
build a very high-quality example. The
vocoder’s primary claim to fame is where
human speech is superimposed on a
musical instrument. A popular example
is the creepy 1961 recording Sparky’s
Magic Piano – see https://youtu.be/
Km19Iohd1YA (strictly speaking, Sparky’s
Piano used a Sonovox which was not a
vocoder, but a pair of special loudspeakers
(compression drivers) applied to the
vocalist’s throat. It was also used in Dumbo
for the talking train whistle.)
A more recent example is the definitive
vocoder song, Laurie Anderson’s 1981
release, O Superman – https://youtu.be/
S39NaDPNDtk (see our editor’s prized,
pristine 12-inch vinyl copy in Fig.1!). It was
recorded with the Roland VP-330: www.
vintagesynth.com/roland/vocoder.php
Origins
As with most electronic audio/music
technology, vocoders were originally
developed for telecommunications. Invented by Homer Dudley at Bell Labs in
Fig.1. Laurie Anderson’s O Superman, a
vocoded treat in avant guard minimalism
using the Roland VP-330.
Analysi s f ilters
( double- tuned)
1936, they enabled more speech channels
to be put through long low-bandwidth
cables and other transmission systems.
Vocoders are an analogue data compression system which operate by representing
speech data in its simplest form, amplitude changes of frequency bands, akin to
a spectrum analyser. These slow changes
can then be transmitted and the speech
resynthesised at the other end. The military used them to disguise voices and
encode secret messages. This is where
the name comes from: VOice enCODER.
A vocoder can be thought of as a system
for transferring the spectral energy of a
speech signal onto another sound. An unfiltered organ or string synthesiser works
especially well. If white noise is used, the
vocoded voice will sound like a whisper.
An analogue vocoder is complex, the
design I’m presenting consists of 28
precision filters and 14 voltage-controlled amplifiers (VCAs). For a ‘normal’
Smoothing
R ectif ier low- pass f ilters
f1
D rive
amplif ier
Multiple channels
C ontrol
vo ltages
to V C A
Speech input
( modulation)
f1 2
I N P U T SE C T I O N
O U T P U T SE C T I O N
Analysi s f ilters
( double- tuned)
f1
V C A
D rive
amplif ier
Mixe r
Multiple channels
Fig.2. Software-based vocoders are
possibly the cheapest entry point for
most musicians, assuming you have
a powerful computer and digital audio
workstation software installed. Ironically,
they can be a good way of optimising an
analogue vocoder design before building.
48
C arrier
( string/ syn th/ paino)
f1 2
+
V ocoder
output
V C A
Fig.3. Analogue Vocoder block diagram, showing analysis, control and resynthesis sections.
Practical Electronics | November | 2021
n
V ocoder channel card
Speech
fL
I N P U T SE C T I O N
f L and f U are the lower/ upper f req uencies
of the double- tunded band- pass f ilter
fU
V - to- I
conve rter
O U T P U T SE C T I O N
Syn th
fL
V C A
fU
O utput
D ual V C A shared
with another channel
L
R
System architecture
Mixe r busses
selectable f or
lef t, right or both
Fig.4. Each vocoder channel consists of these internal building blocks.
electronic engineer this is obviously
something best achieved by digital signal
processing using FFT (Fast Fourier Transform) and DSP hardware and software.
Vocoder software plug-ins are the entry
point for most musicians wanting a vocoder sound. In my case, I played about
with the digital Vocodex Channel Vocoder
in FL Studio (Fig.2) to find the best frequencies and Q values to use. However,
digital vocoders have their own problems.
They need a lot of computing power and
often exhibit considerable time delay
or latency. They all use similar maths
which imparts a peculiar ‘under water’
colouration at low signal levels, which I
find unpleasant. Most smartphones use
vocoding as part of their compression algorithms, so most people will be familiar
with this nasty effect. Software vocoders
brought out a £400 Chinese surface-mount
copy of the Roland, called the VC340. Still,
my design is stereo and it’s modifiable –
important features, because having one’s
own unique sound signature is an essential
attribute for electronic musicians.
are also complex to use with too many
features and parameters to be set up. On
the other hand, analogue vocoders have
an immediate playability that lends them
to composition and live performance.
DIY analogue vocoders are expensive
to build, typically costing a few hundred
pounds. However, buying a ready-made
one, such as the Roland VP-330 Vocoder
Plus, or Tim Orr’s Electronic Music Studios
(EMS), will set you back several thousand
pounds: https://bit.ly/pe-nov21-voc
We are getting to the point where analogue electronic musical instruments are
becoming expensive antiquities in their
own right. This is one of the few occasions
in electronics where building ones own can
save serious money. Ironically, just as we
go to press, and after spending five years
building this Vocoder, Behringer have
The input part of a vocoder is concerned
with analysing the changes of a signal,
usually speech in musical applications,
and often called the ‘modulator’ in communications. This section generates
slowly changing DC control voltages
representing the amplitudes of the different frequency bands.
The output part is concerned with the
generation or synthesis of the final sound.
The block diagram is shown in Fig.3. In
music applications, chords, from say a piano or string synthesiser, (or the carrier/
excitation in comms) are split into corresponding frequency bands. These bands
are then modulated by the control signals
derived from the input section. This requires a VCA to control the output of each
frequency band. The end result is that
the piano or other musical input can be
made to ‘talk in tune’. Alternatively, the
control signals can be patched from one
frequency band to another to make crazy
noises and facilitate speech scrambling.
The internals of the frequency-band
modules are shown in the block diagram
Fig.5. Early photo of the prototype analogue vocoder. No knobs yet.
Practical Electronics | November | 2021
49
f L ( – 3dB) = 4 4 6 H z
fL = 4 7 2 H z
f U = 52 9 H z
f U ( – 3dB) = 56 0 H z
f c=
50 0 H z
0 dB
=
Bandwidth = 1 / 3 octave
( typ ical f or vo coder)
Single Q = 4 . 38
f ilter at 50 0 H z
fL
– 3dB
Q
fU
= f c / ( f U ( – 3dB) – f L ( – 3dB) )
= 50 0 / ( 56 0 – 4 4 6 ) = 4 . 38
– 3dB
Fig.6. Prototype Vocoder channel PCB. Each channel frequency has to have different
values of filter capacitors. (The final version will be double-sided to remove all those links.)
and are further expanded in Fig.4. Each of
these are made into a PCB card that can
be plugged into a bus-board to connect
them all together. The prototype Analogue Vocoder is shown in Fig.5.
This design uses 12 band-pass filter
modules with frequencies optimised for
speech; plus, the system topped and tailed
by fourth-order low-pass (fc = 120Hz) and
high-pass filters (fc = 8kHz). This builds
to a total of 14 filters, all fourth-order,
and requiring 112 close-tolerance capacitors which represent a major chunk of
the parts bill. The EMS vocoder has 22
band-pass filters! A prototype band-pass
vocoder channel board is shown in Fig.6.
Band-pass filters
There are many ways to build a band-pass
filter. The state-variable and its alternative the bi-quad are very effective. The
bi-quad’s resonant frequency (fc) can be
adjusted with just a single resistor; and
the state-variable’s fc with two resistors
– see Fig.7. However, both these configurations need three op amps per filter,
which would be excessive in a multiple
filter system like a vocoder. For perfect
frequency band discrimination, digital
brick-wall band-pass filters are used –
but I suspect in musical applications
these would sound horrible because of
the group-delay-induced ringing.
Ω
Ω
Y
Better bring in Q, Bond
The band-pass frequencies in the speech
band need to be spaced at around one
third of an octave steps. This necessitates
a filter Q of around 4.38, to put the cutoff (−3dB) frequencies in the right place.
To obtain a flatter overall frequency response, double-tuned band-pass filters
can be used, which comprise two filters
whose curves overlap, as shown in Fig.8.
These have two resonances close together with a Q of 8.8, which gives a ‘flat’
top with initial steep (−40dB per octave)
4 . 3nF
R 2
R
–
+
R
Fig.8. By putting two high Q band-pass
filters in series a good compromise curve
can be obtained with steep skirts and a
flat top. This forms a double-tuned bandpass filter, familiar to most radio designers.
Ω
–
I nput
R
–
+
R
Bandpass output
1
1
=
R
R
–
+
O ptional input
attenuator
resistor
f
1
√R 1 R 2
=
Y
1
R
–
Ω
–
+
+
1
1
R 2
2
R 1
2
2
Ω
I nput
O utput
0 V
X
R
Ω
4 . 3nF
Ω
+
Ω
X
I nput
R 1
Ω
f
Ω
f L and f U are the
lower/ upper f req uencies
of the double- tuned
band- pass f ilter
f
Altering this
resistor
changes f c
R
1
=
–
a lot! )
1
√R 1 R 2
=
f
1
= 4 . 4 2 nF
+
Fig.7. The state-variable (above) and bi-quad (below) are the ideal building blocks for
band-pass filters but are too complex and expensive for the multiple channels needed
in a vocoder. However, for a vocoder with a few channels where the parameters need
front panel controls, the bi-quad would be used.
Fig.9. The multiple-feedback filter, the
simplest and cheapest. High gain makes
it noisy. Two of these are needed to make
the flat-topped filter. An input attenuation
resistor to ground is often added, but the
parallel resistance of the attenuator must still
equal 3.6kΩ to avoid altering the response.
50
Practical Electronics | November | 2021
f
=
/ R
Bandpass output
Q
86 0 H z,
= 1 0
Q
R 3
I nput
R 1
C 1
33nF
Ω
R 2
C 4
33nF
Ω
R 6
R 4
–
+
Ω
R 5
= 1 0
C 2
33nF
Ω
–
R 1
+
Ω
Fig.10. A double-tuned band-pass filter circuit. One-third octave bandwidth with a
centre frequency of 950Hz.
8kH
C f
I nput 2 . 2 nF
R f
C f
2 . 2 nF
z high- pass f ilter
Ω
Ω
2 . 2 nF
2 . 2 nF
+
IC 1 a
–
R f
F or speech analysi s
f ilters use T L 0 82 , and
f or the output section
use low- noise
N E 5532 devi ces.
O utput
IC 1 b
Ω
–
Ω
O p amp selection
Ω
Ω
Ω
2 7 0 pF
Ω
2 7 0 pF
0 V
1 2 0 H z low- pass f ilter
R f
1 µ F
I nput
Ω
Ω
R f
Ω
C f
2 2 0 nF
C f
2 2 0 nF
2 2 0 nF
+
Ω
Ω
IC 2 b
–
+
IC 2 b
2 2 0 nF
Ω
–
Ω
Ω
Ω
Ω
O utput
C 2
1 0 0 nF
0 V
=
1
fC C
L et f C = 1 kH
z
Ω
= 1 0 0 nF
C
P eak gain = 2 f or this f ilter
Ω
C 1 - C 2 = 1 0 0 nF
Ω
F or a typ ical vo coder Q
of 5:
Ω
Ω
Fig.11. The dual-amplifier band-pass
configuration filter. Twice as many op
amps, but more controlled gain.
dual-amplifier band-pass configuration
(DABP) developed by Sedra and Espinoza, and shown in Fig.11. This avoids
the excessive noise gain problem of the
multiple-feedback type. However, it adds
24 extra op amp sections to the vocoder,
so we’ll leave this for a future upgrade.
High / low-pass filters
The high-pass and low-pass filter topology
is basically the same as the vocoder channel topology (Fig.4), but with
the band-pass filters replaced
with fourth-order high-pass
and low-pass filters. These filter
circuits are shown in Fig.12.
These extra modules are
only needed if a full frequency response is desired for a
stand-alone-performance vocoder. In a band or studio
situation where there is a bass
player and drummer for example, the vocoder only needs to
cover the speech band – ie, the
middle frequencies of around
100Hz to 8kHz. A treble-boosted portion of direct speech
signal can be mixed in to give
O utput
a clean top-end. By eliminating
the high-pass and low-pass vocoder modules one saves £70,
and gains a less cluttered mix
into the bargain.
0 V
N ote: same P C B can be used f or both circuits, j ust swap R f and C f positions to make low- pass or high- pass f ilter card
Fig.12. For the top and bottom bands of a full-range vocoder, fourth-order low-pass and high-pass
filters are used. These require a different PCB and are not needed for a speech-only vocoder.
Practical Electronics | November | 2021
R 3
+
R
simple design is that the gain is 50 at a
Q of 5, which is a problem because the
op amp output will clip.
For third-octave double-tuned filters,
the Q of each section needs to be 8.8, the
practical limit for this type of filter. A circuit of a double-tuned multiple feedback
filter is shown in Fig.10. The signal has
to be attenuated on the input to the bandpass filters by even more than for a single
filter. Therefore, the signal-to-noise ratio
is poor. This has little effect in speech filters, since these are followed by full-wave
rectifiers to produce smoothed DC control
voltages for the VCAs. On the carrier/output section it is more significant, but it is
only the output filters after the VCAs that
contribute continuous noise. The noise
from the filter stages preceding the VCAs
is effectively muted when the VCAs cut off.
An alternative filter, which I intend
to try at some point in the future, is the
+
–
Ω
D C path
to ground
req uired
0 V
The circuit shown in Fig.9 is the simplest
band-pass filter; it’s called the multiple-feedback filter because it uses two
feedback paths. The disadvantage of this
R 2
Ω
I nput
C entre f req uency of whole ( f lat- topped) f ilter
is 9 50 H z, total Q of f ilter is 5.
Multiple-feedback
band-pass filter
0 V
O utput
0 V
skirts. The result is much less peaky than
a single filter and the out-of-band attenuation is double. The frequencies of the
two filters are selected so as to correspond
with what would be the −3dB points of
a single filter. So for a 500Hz (centre frequency) double-tuned filter, the resonant
frequencies (fL and fU) of the two filters
would be 472HZ and 429Hz respectively.
Note the slight asymmetry is due to the
neeed to match (approximately) exponential frequency scaling. The double-tuned
technique was originally developed for
radio intermediate frequency transformers in superhet radios.
C 1
1 0 0 nF
Ω
C 3
33nF
Ω
R 5
Ω
R 4
–
z,
+
1 . 2 5kH
Next month
That’s all for this month – in
the next issue we’ll start to look
in more depth at the thinking
behind the design.
51
