PAPERS

Effect Design*
Part 1: Reverberator and Other Filters

JON DATTORRO, AES Member

CCRMA, Stanford University, Stanford, CA, USA

The paper is a tutorial intended to serve as a reference in the field of digital audio effects
in the electronic music industry for those who are new to this specialization of digital signal
processing. The effects presented are those that are demanded most often, hence they will
serve as a good toolbox. The algorithms chosen are of such a fundamental nature that they
will find application ubiquitously and often.

0 INTRODUCTION deals with the impact of finite precision. That will de-
mand consideration when someone complains of too
This paper is intended to serve as a point of reference much noise or grit in your output signal.
in the field of digital audio effects for the electronic The two most asked-for effects are chorus and rever-
music industry. It is for those who are new to this special- beration. Reverberation creates an ambient space in the
ization of digital signal processing so as to advance their perception of the listener. The reverberator presented
skill level at inception. The effects presented herein are herein is the smallest recursive network we found that
those demanded most often; hence they will serve as a meets subjective requirements of good sounding rever-
good toolbox. The algorithms chosen are of a fundamen- beration. This reverberator is not analyzed in great math-
tal nature and therefore will find application ubiquitously ematical detail; it is best explored by tinkering, because
and often. They include one for reverberation, two for that is how it was developed. There are few enough
filtering, two for delay-line interpolation, one for chorus knobs so that the sonic impact of each is readily dis-
(as well as vibrato and flanging), four for sinusoidal cernible.
oscillation, and one for noise generation. Filtering for musical purposes involves somewhat dif-
It is not necessary to start reading the paper from ferent considerations than what is taught in standard
the beginning. The overall tone of the paper is tutorial, texts on digital signal processing (DSP). The most nota-
stressing concepts. The supporting mathematics go to ble departure is that of the half-power excursion 1 of the
some depth in those cases where the algorithms are ana- magnitude response when regarding audio filters that are
lyzable. The reader is not required to delve that deeply; typically shallow. Simple and accurate design equations
in some cases knowledge of the results alone is suffi- for an easy-to-operate second-order notch filter and reso-
cient. The mathematics serve to develop concepts, to nator are developed from the musician's point of view.
justify conclusions rigorously, and to offer aid when one A unifying framework for both filter types develops into
runs into trouble. Of course, the best way to learn is to the R e g a l i a - M i t r a topology, which facilitates paramet-
try the algorithms and invent one's own. ric equalization. We then apply the same simplifying
Our hardware reference standard is a dedicated 24-bit concepts to the musician's popular second-order all-pole
two's complement fixed-point digital signal processor filter, which is used for a wide range of purposes, span-
chip [1], typically having 48-bit accumulation of prod- ning wa-wa to dynamic noise rejection. The musical
ucts, but these algorithms will certainly run on any per- filtering sections culminate with a unique realization of
sonal computer. Nevertheless, much of the mathematics that popular filter--the versatile and quiet Chamberlin
filter topology, the digital analogue to the Moog voltage-
* Manuscript received 1996 March 14; revised 1996 Sep-
tember 14 and 1997 June 28. 1 A relative as opposed to absolute measure.

660 J. Audio Eng. Soc, Vol. 45, No 9, 1997 September

PAPERS EFFECT DESIGN

controlled filter. tions of reverberators rely on all-pass circuits embedded

We scrutinize linear interpolation as a means for delay within very large globally recursive networks. The all-
modulation. The modulating delay line forms the basis pass circuits themselves have recursive delays measured
of many standard audio effects. The inherent filtering in hundreds of sample periods, whereas the cumulative
artifact of the linear interpolation process is often over- delay around a large recursive network can total on the
looked, however. We offer an alternative, called all- order of tens of thousands of samples.
pass interpolation, which avoids the pitfalls in some The early successful commercial inventors of these
circumstances and sounds very analog. The chorus effect complicated networks were Griesinger and Blesser. Un-
is well served by this alternative method of interpolation. fortunately they have written little on this topic. Moorer
Chorusing emulates a multiplicity of nearly identical and Gardner have turned the art more into a science.
sound sources. When only two sources are emulated Moorer elevates the seminal but crude work of Schroeder.
(two voices, including the original), we consider that to Gardner provides a technical chronicle of developments
be the industry-standard chorus effect. This perceptually in the art of reverberator design, where he also furnishes
pleasant effect is hard to describe and must be experi- a synopsis of his complete translation of the French van-
enced to be fully apprehended. guard Jot. 2 We do not provide sufficient background
Sinusoidal oscillators are found within nearly every material to permit the reader to fully understand the
audio effect. Although oscillators can generate sound, development of the reverberation network for plate emu-
more often than not they are used to control some modu- lation presented herein. The reader is encouraged to refer
lation process. Delay modulation is a key to successful to the references [2]-[9, pp. 1-28].
reverberator design. Writing a few simple instructions,
it is easier to design a terse algorithm to generate a sine 1.1 S i m p l e R e v e r b e r a t i o n Network 3
wave than it is to employ a table lookup. The algorithmic Fig. 1 shows one particular network for producing
approach also results in a purer sinusoid. We examine reverberation. We like this topology for several reasons:
several efficient methods of sinusoid generation, and we 1) It has simple knobs, which easily control particular
offer guidelines to aid in the choice. aspects of the reverberated sound, such as input and decay
Noise generation, seemingly the antithesis of sinusoid diffusion (decorrelation), decay rate, high-frequency
generation, is discussed. The exceedingly simple maXimal- damping, and input signal bandwidth. 2) The style of
length pseudorandom noise generator is presented as a the topology is more computationally efficient than most
pleasant and soothing sound source. Not only does this others known. 3) It has demonstrated applicability to a
simple circuit produce a pseudorandom bit stream, it broad range of signal sources.
also emits a sequence of pseudorandom multibit words, We selected the network in Fig. 1 for presentation
each repeating only once per cycle. The cycle time can because it is the smallest reverberation network we found
easily be designed to exceed the duration over which (in memory and complexity) that is good sounding. We
the human ear can identify patterns. The classical litera- believe that there must be a limitless variety of such
ture on these circuits demonstrates that the autocovari- networks, however. The question naturally arises as to
ance of the single-bit pseudonoise sequence is a Kro- why the simple digital network shown produces such
necker delta. Hence the single-bit noise is uncorrelated convincing reverberation. We can answer this only
and spectrally white. In the multibit case we find that qualitatively.
to be only approximately true. Short-lived exponential Consider the plucked string of a violin. Its envelope
patterns are visible within the pseudonoise sequence, may be described as having a coherent exponential de-
revealing correlation. Hence the power spectrum of a cay. It is this character that is theorized to be one of the
multibit noise realization via this circuit cannot be per- primary discriminants of nonreverberated sound. Rever-
fectly white without equalization. We show how the berating this sound, on the other hand, would tend to
multibit pseudonoise sequence can be precisely modeled randomize the string envelope and phase, producing a
as linear FIR filtering of the single-bit sequence. Thus bumpier, extended, more diffuse and dynamic decay.
the power spectrum of the uniform amplitude-distribution This oversimplified qualitative description of the pro-
multibit sequence is known, and we suggest a simple cess of reverberation has actually found its way into
method of equalization. early commercial products. Long before DSP chips
could be integrated into sampler type synthesizers, re-
verberated sampled sound was simulated by altering the
1 REVERBERATION
decay characteristics of recorded dry samples by ran-
Digital reverberators are like paintings. There are zil- domizing an overlaid envelope applied at playback.
lions of them, all of different colors, as no one wants While not absolutely convincing, this kind of aural cue
the same painting in every room. The engineer's pipe
dream of the universal reverberator may never be real-
2 Jot recently formulated an analytical design method for
ized. A treatise on artificial reverberation would easily recursive reverberators. The work is based on a unitary (loss-
fill volumes. In the past, these networks were so difficult less) feedback loop in a state-space network, where he claims
to analyze (like Bach fugues) that they have traditionally arbitrary time and frequency density
3 This discussion is adapted from conversations with Barry
been invented through experimentation. The reason for Blesser and David Griesinger and is supplemented by Appen-
the difficulty is that even the most efficient implementa- dix 1 in Section 1.5

J. Audio Eng Soc., Vol. 45, No. 9, 1997 September 661

DATTORRO PAPERS

was enough to cause pioneers [ 3 ] - [ 5 ] to question the hertz, It can also be theorized that the limit on the num-
premise of Schroeder's precipitative work with delay ber of achievable eigentones is proportional to the total
lines at Bell Labs during the early 1960s. delay-line memory [4]. From our current perspective we
One can deduce from Schroeder's work [7] that to
achieve the ideal of colorless reverberation, the e i g e n - 4 An eigentone of a network in this context is a circuit
t o n e 4 density of the network needs to approach 3 per resonance.

bandwidth

z'~ I
predelay
xR

1. - bandwidth

,nputdiffusi1 on / / 7nputdi"usion
I

--r--~ J ~ ~
in0ut0,.os,on ,o0 ,0,..sion

672 + EXCURSION ] 908+ EXCURSION J

note sign _ ,j .~.

decay diffusion 1 ' diffusion 1

1. - damping

damping
decay

;at

z.3 o |
decay diffusion 2

Fig. 1. Simplified plate-class reverberation topology in the style of Griesinger. For output tap structure (YL, YR) see Table 2.
Delay-line taps at nodes 24 and 48 are modulating.

662 J. Audio Eng, Soc, Vol 45, No. 9, 1997 September

PAPERS EFFECT DESIGN

know that emulation of physical spaces can be convinc- fusers (the lattices) followed by another set of four tank
ingly performed using sample rates as low as 2 0 - 2 4 diffusers, the latter arranged so as to feed back on them-
kHz. This is true because of typically rapid acoustical selves globally. The first set of diffusers acts to quickly
absorption in the high-frequency region, and because decorrelate the incoming sound somewhat, preparing
the desired output is a mix with the dry input signal. This that sound to be looped indefinitely in the holding tank
bandwidth would then require about 30 000 eigentones, formed by the second set of diffusers. What we hear
hence about 64K words of delay-line memory. In the comes from a large set of output taps (not shown) located
1960s, that amount was not economical. 5 within the tank.
In reverberator design, while a good general rule re-
garding delay-line memory is certainly "the more the 1.3.1 Input Diffusers
better" [4], the efficient reverberation network shown in All the diffusers are all-pass filters having the topol-
Fig. 1 stands as a testimony 6 that Schroeder's eigentone ogy of a lattice. The purpose of the four input diffusers
density criterion, predicting about 88K words of mem- is to decorrelate the incoming signal quickly before it
ory, is not a hard and fast rule. Of at least equal impor- reaches the tank. The tank recirculation can sometimes
tance are the decorrelation of the decay and the associ- become perceptible as strong cyclic events if the input
ated time density of the echoes, that is, one must achieve signal is not preconditioned in this manner. This function
a balance between eigentone density and echo density. becomes especially important for the successful rever-
beration of percussive sounds. One may think of this
1.2 C o l o r function as signal-phase randomization, to reduce pea-
On the other hand, our reverberation network's signal kedness and other strong features of the input waveform.
response is not colorless. Empirically we find that some No diffusion corresponds to zero-valued all-pass coef-
of the most sought after commercial reverberators are ficients, while coefficient magnitudes close to unity pro-
somewhat colored in their frequency responses. This duce buzzing that is local to the afflicted all-pass filter.
means that their outputs impose some conspicuous audible Optimum diffusion for the all-pass filter lies somewhere
resonances upon the input signal. Consequently it is not in a region closer to 10.51 than to the extreme values of
unusual to find as many musicians and recording engineers the coefficients. The preset values given in Table 1 were
who like a particular reverberator as those who do not. determined by trial and error.
We also find that some recording engineers do n o t
want an accurate emulation of a physical space, because 1.3.2 Tank
the reflection density takes too long to build. Instead, We identify the reverberation tank as the recirculating
they sometimes want instantaneous high-density reflec- four lowest diffusers in Fig. 1. We call it a tank because
tions with smooth exponential decay of the envelope, its purpose is to trap the incoming sound by making it
having randomization in only the phase trail. This desire recirculate through the global figure eight. The four de-
most closely describes the plate class of reverberators, cay coefficients determine the rate of decay. When the
which we present here. decay coefficients are set very close to 1.0 (and the
damping filter within the tank is turned off), the sound
1.3 D i s c u s s i o n of t h e R e v e r b e r a t o r will remain held in the tank indefinitely. That in itself
Scrutinizing the reverberation topology in Fig. l, we is a neat effect, but unless the sound metamorphoses
can break it down into a cascade set of four input dif- while in the tank, it is easy for us to detect the looping
pattern of sound. The purpose of the diffusers within
the tank, then, is to eliminate any aural pattern in the
5 The Lexicon model 224 digital reverberation system intro-
duced in 1979 originally possessed only 16K words of mem- recirculation. The tank diffusers are not always success-
ory, operating at a sample rate of 20 kHz. That memory amount ful (being signal dependent), and their settings are criti-
doubled shortly thereafter. The Elecktromesstechnik Wilhelm cal to achieve an overall exponential decay; everything
Franz KG EMT-250 digital reverberator distributed in the
United States by Gotham Audio Corporation beginning in must be set by ear.
1977, operated at a sample rate of 32 kHz, having only 8K The tank, in summary, is a simple device whose pur-
words of memory. The precursor to this machine is described pose it is to alter the tail of a decaying sound, as men-
in [10, ch. 2].
6 Given a 30-kHz sample rate and having only 22K words tioned already. The tank diffusers have been further
of memory (not including predelay). grouped into pairs labeled by the knobs "decay diffusion

Table 1. Reverberation parameters default.

Sample rate F s = 29761 Hz

EXCURSION = 16 Maximum peak sample excursion of delay modulation
decay = 0.50 Rate of decay
decay diffusion 1 = 0.70 Controls density of tail
decay diffusion 2 = 0.50 Decorrelates tank signals; decay diffusion 2 = decay + 0.15, floor = 0.25, ceiling = 0.50
input diffusion 1 = 0.750 Decorrelates incoming signal
input diffusion 2 = 0.625
bandwidth = 0.9995 High-frequency attenuation on input; full bandwidth = 0.9999999
damping = 0.0005 High-frequency damping; no damping = 0.0

J. AudioEng. Soc, Vol. 45, No. 9, 1997 September 663

DATTORRO PAPERS

1" and "decay diffusion 2." The tank diffusers have spontaneous tones can be eliminated through the use of
overlapping functionality. The dichotomy we make is magnitude truncation (truncation toward zero; see Part
aurally subtle and pertains to the temporal location of 3, Section 9.2, Appendix 8) of the double-precision in-
the diffusers in the tank with respect to a stereo tank termediate results written out to single or lower precision
input, that is, exactly w h e n they diffuse the tank signal delay-line memory. Magnitude truncation is well k n o w n
with respect to the signal onset. The effect of these to subdue limit cycles in digital networks composed of
knobs is best observed using a percussive input, or what ladders and lattices [9].
Griesinger refers to as a "pink click.'7 Only the recursive circuits require magnitude trunca-
tion. In Fig. 1 the write to the predelay does not require
1.3.3 All-Pass Lattice Topology magnitude truncation. If delay-line memory is 24 bits
Each diffuser has been given the topology of a two- in width, then the need for magnitude truncation is obvi-
multiplier lattice. The eight lattices shown in the rever- ously lessened when compared to having delay-line
berator schematic in Fig. 1 are used in this reverberation memory of only 16 bits in width.
effect as all-pass filters, each having a long impulse Magnitude truncation, in the specific case of reverber-
response time.8 The two coefficients within each individ- ator tank topologies employing lattice or ladder all-pass
ual lattice must remain identical to maintain the all- circuits, can reduce the network noise floor by 12-24
pass transfer function, which is insensitive to coefficient dB after the input signal is removed. The reason this is
quantization. The recommended range of these coeffi- true is that the predominant noise mechanism is zero-
cients is from 0.0 to 0.9999999 (q23; see Part 3, Section input limit-cycle oscillation,11 a multiplicity of which is
9.1, Appendix 7) If the lattice coefficients should exceed perceived as a whooshing ocean noise floor. The magni-
1.0, instability would result. Making them both negative tude truncation makes the reverberator output eventually
will change the character of the impulse response 9 but go to absolute zero, two's complement. The disadvan-
does not destroy the all-pass transfer. This change in tage to its use is that the THD + N (total harmonic distor-
character is exploited in the lattices having the coeffi- tion + noise) of a steady-state sinusoid through the
cients called "decay diffusion 1" in the schematic. This linear reverberator network can be increased by any-
character change further enhances the dichotomy be- where from 0 to 6 dB.
tween the two pairs of tank diffusers.
All-pass response is the forced (steady-state) response 1.3.5 First-Order Filters
of each lattice output with respect to its own input.l~ The three single-pole low-pass filters used for input
Because the impulse response of each individual lattice signal bandwidth control and reverberator tank damping
within the reverberator schematic is so long, in some will not clip prematurely at any node [11, ch. 11.3],
cases the integration time constant of the human hearing [12, p. 857] when implemented as direct form I. The
system is exceeded. This means that an all-pass filter damping filters cause high frequencies to decay within
output may be perceived as discretized events, that is, the tank more quickly than low frequencies. On the
not all pass. input-bandwidth filter, the bandwidth coefficient tracks
This all-pass lattice topology tends to clip prematurely the cutoff frequency. In contrast, the damping coefficient
at internal nodes, so the input to each lattice cannot be is high when the damping filter cutoff frequency is low.
presented with a full-scale signal at all frequencies. We The recommended range of these coefficients is from 0.0
like this all-pass lattice, however, because it is efficient to 0.9999999 (q23; see Part 3, Section 9.1, Appendix 7).
in its implementation. Because they are all first-order low-pass filters, any
low-level zero-input limit cycles they might produce
1.3.4 Magnitude Truncation would be at dc, that is, they will not produce tones like
Lattices produce distinct low-level tones, after the the lattices [11, ch. I 1.5]. Any signal-truncation noise
input signal has been removed, known as zero-input power spectrum generated by the filters themselves will
limit cycles. The origin of these tones stems from ongo- be centered at de, since it follows the pole frequency.
ing signal quantization in a recursive topology. The The peak gain of the noise power spectrum is not great
because typically the lone pole is relatively far from the
unit circle. 12
7 A click source having a pink spectrum.
s The impulse response is that of an upsampled first-order 1.3.6 Output Tap Points
all-pass filter. This filter basically has an exponentially de-
caying impulse response with or without a multiplicative factor From the pseudocode note that the delay-line tap
of ( - 1)n - 1 , depending on the sign of the coefficient. The up- structure forming the stereo output signal YL and YR is
sampling factor L is determined by the number of samples in
the lone delay line z -L within the lattice. The up-sampling an all wet (reverberated) signal (Table 2). This particular
process inserts L - 1 zeros between every sample of the output tap structure is characteristic of the plate emula-
impulse response of the corresponding first-order all pass filter.
9 Via the multiplicative factor ( - 1 ) "-~ on the impulse
response.
~0 The reverberation network as a whole does not have an II Here we use the term "limit cycle" in the classical DSP
all-pass transfer function, although we would like that to be sense.
the case. Smith [9, pp. 1-28] has found a way to make an 12 If instead the filters were high pass, limit-cycle tones
entire reverberation network all pass. Smith's method is based might be produced at Nyquist while the truncation noise power
on the interconnection of lossless waveguides. spectrum would also be concentrated there.

664 J. Audio Eng Soc., Vol. 45, No 9, 1997 September

PAPERS EFFECTDESIGN

don class of reverberation networks. 13 Also note that the sets, the modulation is a godsend. In the case of piano,
output tap structure produces a synthetic stereo image the modulation, though slight, may be objectionable be-
because the stereo input is converted to a monophonic cause of a perceived vibrato.
signal at the reverberator input for this particular topol- Ideally, all the delay lines in the tank diffusers should
ogy. Normally, the desired output is a mix of the stereo be modulated using different modulation rates and
reverberated signal YL and YR with the original (dry, full- depths. In that case, the diffusion burden becomes more
bandwidth) stereo input signal x L and XR. distributed. Hence the required rate and depth of modu-
lation are lessened for each diffuser. When computation
1.3.7 Delay Modulation time is a constraint, then one should preferentially select
Linear interpolation or, better yet, all-pass interpola- the stereo pair of the diffusers appearing earliest in the
tion (as discussed in Part 2, Section 5) can be efficiently tank, as we have, to maximize the increase of effective
employed to modulate slowly x4 the nominal tap point of resonances. In this case, the same rate and depth are
the two indicated delay lines in the schematic. A slight used for each diffuser in the pair, but we use a quadrature
modulation will introduce undulating pitch change into oscillator to decrease the correlation. (Sinusoidal oscil-
the tank. For signals with much high-frequency content, lators are discussed in Section 7.) The differing delay-
such as drum sets, these built-in modulators serve to line lengths of all the diffusers also serve to decrease
break up some pretty audible modes, that is, the amount the correlation.
of tank diffusion is effectively increased. As explained in Section 4; linear interpolation for de-
Barring air currents and temperature fluctuations, lay modulation will introduce time-varying low-pass fil-
there is no analogue to this modulation process in a tering as an artifact, thus supplying some unaccounted
real room (unless the walls are moving). Without the damping to the tank. All-pass interpolation overcomes
modulation, we may well describe the imaginary space this particular problem and is perfectly applicable to
emulated by the given digital network as being enclosed reverberators because the required pitch change is
by a picket fence. The slow modulation serves to increase microtonal. 15
effectively the sheer number of resonances (eigen-
tones, modes of oscillation, picket density) in the tank. 1.4 C o n c l u s i o n
The number of resonances in a real room, hall, or plate Choosing a particular reverberator for a particular ap-
is probably far beyond what is existent in our little (non- plication is commonplace, and purveyors of such equip-
modulating) reverberation network. In the case of drum ment have been known to purchase an audio signal pro-
cessing box just to acquire one particular algorithm.16
At some level, the choice of reverberator becomes a
la The physical "plate," actually resident in some contempo-
rary recording studios, fills a small room in some embodi- matter of taste, much like art. There is no one universal
ments. The best plates are constructed using a solid g.old foil. reverberation network that satisfies everyone for each
The input signal is typically injected onto the plate via one or and every application; we speculate that there never
two transducers, while each output is the sum of multifarious
signal taps, each tap transduced at a different location on will be.
the plate.
14 At a rate on the order of 1 Hz, and at a peak excursion 1.5 A p p e n d i x 1: R e v e r b e r a t i o n R e c o l l e c t i o n s
of about 8 samples for a sample rate of about 29.8 kHz.
Dear Jon,
What you wrote was fine, but it stimulated my mem-
ory of additional snippets. Feel free to use what you
Table 2. Output taps. want.
I had a personal conversation with Manfred Schroeder
/********* left output, all wet *********/
in the late 1970s and I asked him the question about
accumulator = 0.6 X node48_541266]
what the phrase "maximal incommensurate" delay val-
accumulator += 0.6 x node48_54[2974]
ues meant, as it appeared in one of his reverberation
papers. His answer was particularly interesting. This is
accumulator -= 0.6 X node55_59[1913]
a paraphrase based on my tired memory:
accumulator += 0.6 X node59_63[1996]
We did the electronic reverberation for amusement
accumulator -= 0.6 X node24_30[1990]
because we thought it would be fun. Since it took the
accumulator -= 0.6 x node31_33[187]
better part of a day to do 10 seconds of reverberation,
YL = accumulator - 0.6 X node33_39[1066]
we only ran one sample of music. The notion of delay
time selections was random in that we just picked a
/********* right output, all wet *********/ bunch of numbers and there was no mathematical ba-
accumulator = 0.6 X node24_30[353]
accumulator += 0.6 X node24_30[3627]
accumulator -= 0.6 X node31_33[1228]
15 The sinusoidal low-frequency oscillator driving the
modulator must have a rate of update that is the same as
accumulator += 0.6 X node33_39[2673] the audio sample rate, that is, the two sample rates must be
accumulator -= 0.6 X node48_54[2111] identical. Otherwise, aliasing artifacts will be introduced
into the audio signal path.
accumulator -= 0.6 X node55_59[335] 16 Much like buying a Compact Disc because one likes
YR = accumulator - 0.6 X node59_63[121] the title track.

J. AudioEng See., Vol 45, No 9, 1997 September 665

DATTORRO PAPERS

sis. We just wanted to prove it could be done. that is proportional to the square of frequency. All elec-
He never related this work to his more profound math- tronic simulations tend to have a constant density. The
ematical and perceptual research, specifically the work reason is that in a three-dimensional space, the speed of
on the required 3-eigentone/Hz density and the frequency- sound along a dimension is proportional to the sine of the
phase statistics in a random physical space. wave front direction, whereas in an electronic structure it
The original EMT reverberator, model 250, operating is always constant.
at a 32-kHz sample rate, used a main memory of 8K That is what I remember, so do what you wish with
words, and the required eigentone density was emulated it. Best of luck.
entirely by randomizing delay lines. Another interesting Sincerely yours,
fact is that colorless reverberation, using all-pass struc- BARRY BLESSER
tures, is perceptually not colorless. Even white noise Blesser Associates
passed through an all pass will not sound like real white Electronics & Software Consultants
noise. When passed through many such all-pass struc- Belmont, MA 02178, USA
tures, it in fact sounds like a machine shop rather than
random noise. It still measures spectrally flat. The reason 2 MUSICAL FILTERING
is that frequency regions get bunched in time. It is very
much like a chirped sine wave in radar having a purely The first-order recursive filter is by far the safest and
fiat spectrum but being very different from white noise. most economical choice. Low in noise, it should be used
The second- and higher order statistical terms out of an wherever possible, and in cascade if necessary. For the
all pass are very, very different from a real random design of shelving filters, which are conventionally first
process. The utility of an all pass is to pass all frequen- order, refer to [ 13]. When a filter having a steeper transi-
cies through so that each all pass can see the same spec- tion band is desired, it is usually sufficient to employ a
tral density, otherwise comb peaks would align and dom- second-order filter. Musical filters do not often see or-
inate. Parallel structures of non-all-pass elements ders higher than that. 17
achieve a similar issue in that each structure gets fed In this section we discuss filtering requirements for
the full spectrum. All-pass elements are more critical musicians whose criteria are quite different from those
for small delay values. An all pass within a larger loop of the electronics engineer. Our treatment of filtering
must be used with great care since it has a sinelike will consider only the second-order case and predomi-
variation in group delay. Hence the effective loop time nantly all-pass topologies. The applications of these fil-
and reverberation time vary with frequency. After many ters are broad; we note a particular suitability to paramet-
trips around the loop, the result will be very colored. ric equalization. The more involved topic of truncation
Schroeder's had several analyses about reverberation, noise recirculation is not discussed in this section, al-
but his 3-eigentone/Hz theory, which maps to 3 seconds though we do discuss limit cycles and internal signal
of memory, can be looked at in many ways. His result overflow. The more curious reader is referred to [12] to
was empirical, based on listening tests. Consider two find remedies for truncation noise within the direct form
eigentones, or poles, separated by 1 Hz and located in I topology.
the s plane with a real part of - 10 Hz. When excited, For those readers new to digital filtering, the eminent
this will produce two damped exponentially decaying theoretician, practitioner, and mentor of DSP and elec-
frequencies which differ by 1 Hz. Hence there will be tronic music, Julius O. Smith, presents a splendid intro-
a 1-Hz envelope beat, which is clearly audible. Now duction to classical digital filter theory in [15, ch. 2],
add other eigentones, randomly spaced but still at a requiring only basic mathematical skills. Strawn's audio
distance of - 10 Hz. Assume 10 such eigentones. All signal processing book [15] is written from the musi-
o f them will beat with each other, producing a random cian's standpoint, hence it is highly recommended.
envelope with a spectrum that is crudely flat from 0 to
10 Hz. One can do this simulation in closed form with 2.1 Filter (O) Selectivity
variable excitation of each eigentone. Schroeder's result Electronics engineers are accustomed to think of digi-
actually depends on the nominal reverberation time since tal filters analytically in terms of p o l e - z e r o constellation
that determines how many eigentones will get excited and locus, cutoff frequency, passband ripple, transition
by a narrow-band input. In the early reverberation boxes, band or slope, stopband attenuation, and so on. Musi-
with only 150 ms of reverberation, typically only a few cians and recording engineers are more comfortable
tones would be excited. The envelope had a clear period- thinking in terms of filter parameters--gain or cut, cen-
icity of 6 Hz on average. It sounds bad. Some regions ter frequency, filter Q (selectivity) or bandwidth. For-
had only two eigentones excited with a distance of 2 mally, filter Q is defined as the positive quantity
Hz, which was even worse. Development was much
t% _ t% (1)
more exciting with such limited memory. Today one
Q-Ao~ o~2-COl
can use 1 second of DRAM memory. Many simpler
structures will thus produce good reverberation.
The perceptual simulations deviated from physical 17 When a filter that is steeper than second order is required,
it is advisable to construct it as a cascade of second-order
reality in many ways. For example, a natural three- sections. That will mitigate any coefficient sensitivity or trun-
dimensional space has an increasing eigentone density cation noise problems [ 11, oh. 11.4-11.6], [ 14].

666 J. Audio Eng. Soc., Vol. 45, No. 9, 1997 September

PAPERS EFFECT DESIGN

that is, the center frequency divided by the bandwidth. metrical with the cut filter. We shall shortly see how.
The bandwidth is determined from the particular defini- This is the reason why many of the numbers are exactly
tion of the cutoff frequencies oh and to2 (in radians). the same in both Figs. 3 and 2.
Traditionally the cutoff frequency coincides with an ab- For the resonator (rather, the normalized boost filter)
solute half-power level. In the archetypal case of a steep we acquire the two musical cutoff frequencies % and
unity-gain (0-dB) low-pass filter we recall this level as % , solving the slightly different equation
corresponding to the frequency at which the magnitude-
squared response reaches - 3 . 0 1 dB [ = 10 logx0(1/2)]. 1 -IHb.o=(eJ')l 2 1
But shallow audio filters may not have an absolute half-
(3)
1 -lHboo= (• 1)1 = 2"
power level. So we must refine the definition of cutoff
frequency in terms of half-power e x c u r s i o n , n o t an abso- Like before, the bandwidth is measured halfway u p the
lute level [16]. peak of the magnitude-squared response in Fig. 3. Again
Take, for example, the cut filter magnitude-squared we note that if IHb.o= (- 1)12 = 0 (perfect resonator),
response shown in Fig. 2. This example response has a then the solution to Eq. (3) corresponds to the traditional
Q of 2. We define the two musical cutoff frequencies definition of cutoff frequency.
as corresponding to the level at which Having gained an understanding of musical filter Q,
we begin with two unique and musically useful digital
1 -IHie>)12 1 filter transfer functions, which precisely fit our definition
I - IH~(eJ'%)IZ= 2" (2) of filter selectivity.

We must solve this equation for to. There are two solu- 2.2 Cut Filter
tions, tol and % . Referring to Fig. 2, this equation in- When constructing a notch filter, we expect there to
structs us to measure the bandwidth halfway down the be an absolute zero of transmission at some selected
trough of the m a g n i t u d e - s q u a r e d response. This makes frequency in its transfer function. If we use a filter that
intuitive sense. We cannot use the traditional definition only has zeros (that is, no poles), we can indeed make
of cutoff frequency for this example because the trough a notch. The problem with this approach is that the rest
is not deep enough. But note that if Ino(eJ%12 = 0 (notch of the magnitude response would not be very fiat, as we
filter), then the solution to Eq. (2) would correspond to might like it to be. We might also like a "surgical"
the traditional definition of cutoff frequency. notch, one that has high selectivity. Fig. 4 is an example
The situation is pretty much the same for the resona- showing the magnitude transfer of a badly designed
tor. Whereas the cut filter approaches unity asymptoti- notch filter evaluated along the unit circle in the z plane.
cally at z = • 1, the resonator is loosely defined as a The zero radius is R = 1, whereas the zero angle is
second-order peaked filter having a peak gain normalized 0 = 1 rad. This transfer function has two trivial poles
to unity at its center frequency. The resonator is easily a t z --- 0.
formulated such that its magnitude-squared response is The magnitude response shown in Fig. 4 would pretty
an exact flip of the corresponding cut filter about the much obliterate a musical signal, especially because of
horizontal half-power excursion line, that is, it is sym- the gain at high frequencies. Note that if the zero were

half-power excursion points

IHc(eJ~
(0.769836, 0.834711 ) (1.269836, 0.834711)

I
/ t
bandwidth ~ I -
p o w e r excursion
0.8

0.6 cut depth 2 = 0.669421

Absolute 1 - 0.834711
= 1/2
0.4
1 - 0.669421

mc =2=Q
0.2 1.269836 - 0.769836

tOc=l
i , , i , i , , , , i , , , i l l , i l l
0.5 1 1.5 2.5 3

Fig. 2. Cut filter excursion ~ 1.7 dB.

J Audio Eng. Soc., Vol. 45, No 9, 1997 September 667

DATTORRO PAPERS

m o v e d to a new fixed frequency, the rest of the magni- equation

tude response would change its shape in an undesirable
way. Hence this particular notch filter is not very useful 1 - z =2
for surgical filtering. Hr(z) = (1 - [~) 1 4- ~/(1 "]- [~)z -1 + [~z-2" (5)
In [13] it was shown how to make the passband por-
tions of the notch filter flat, and how to achieve high This filter has a peak gain that is a l w a y s precisely 1,
selectivity. This is accomplished by adding nontrivial regardless of the center frequency. This is characteristic
poles. The result is illustrated in Fig. 5 and expressed of a resonator. The two zeros make the skirts of the
in the transfer function
I1 - 2 R c o s ( 0 ) z -1 + R 2z-21 @ z = e j~
1 + 2~/z - 1 + z - 2
H.(z) = (1 + 13) 1 +~/(1 + 13)z-1 + 13z-2" (4)

2.
This notch filter [Eq. (4)] has an absolute zero having
controllable selectivity there at its center frequency,
while its magnitude at dc and Nyquist is a l w a y s 1, re-
i.
gardless of the center frequency. We must determine
how to obtain a trough of arbitrary depth while main-
taining the other attributes. This would be called a p a r a -
0.5
m e t r i c cut filter. Before we do that, however, we look
at the resonator, which is an exact powerwise flip of this 0.5 1 1.5 2 2.5 3
notch filter about a horizontal. O)
Fig. 4. Poor notch filter; R = 1, 0 = 1.
2.3 Resonator
One use of a perfect resonator in electronic music is IHa(eJ~
to synthesize ping sounds via impulsive excitation. We
1
discuss a more general resonator for use as a musical
Absolute
filter. It is easy to construct a simple resonator using
0.8
only poles. But such an approach has problems similar
to those we encountered with the all-zero notch filter,
0.6
especially with regard to shape, selectivity, and magni-
tude. In particular, the peak magnitude will vary as the
0.4
center frequency is changed to new fixed values.
In [11, ch. 4.3] it was shown how to normalize the
:" O.
height of the resonator peak magnitude as the center
frequency changes, namely, by adding two zeros, one
0.5 1 1.5 2 2.5
at Nyquist and the other at dc. This musically useful to
result is illustrated in Fig. 6 and expressed by the Fig. 5. Notch filter; Q = 2, toc = 1.

IHbnorm(eJ~ 2 half-power excursion points

(0 769836, 0.834711) (1 269836, 0 834711)
1
[ /

\ \ I / / power excursion
0.8

0.6
- skirt depth 2 = 0.669421 .....
Absolute 1 - 0.834711
- 1/2
1 - 0.669421
0.4
s c
=2=Q
1 269836 - 0.769836

0.2

~c = 1

. . . . o15 . . . . { . . . . 115 . . . . '2.5 3

Fig. 3. Resonator excursion ~ 1.7 dB.

668 J Audio Eng Soc., Vol 45, No 9, 1997 September

PAPERS EFFECT DESIGN

magnitude-squared response go to zero at the extremit- Back to the problem at hand, it is easily proven that
ies. When the extremities reach zero, we call this the
perfect resonator [Eq. (5)]. We must determine how to n,(z) - 1 - A(z)
2 (8)
make skirts of arbitrary d e p t h - - t h e resonator. We must
also determine how to place the skirts at absolute magni-
1 + A(z)
tude 1 while achieving arbitrary peak heights; that would H.(z) - 2 (9)
be called a parametric boost filter. We have yet to define
~/and 13. Substituting Eq. (6) into Eqs. (8) and (9), we can derive
Eqs. (5) and (4). This means that we can construct notch
2.4 Musical Filter Topology and perfect resonant filters from an all-pass filter. We
The two transfer functions H,(z) and Hr(z) have some only have left to show that using the all-pass filter we
desirable theoretical and practical properties. First, there can construct cut, resonant, and boost filters as well. We
is a strong bond between Eqs. (4) and (5). Because their will use the fact [Eq. (7)] that at the critical frequency
denominators are identical, there exists one circuit that
can generate both. Second, there is a simple relationship ~ = arccos(-~) (10)
between the coefficients 13 and ~/and the musical filter
parameters toe and Q. the all-pass filter output is 180 ~ out of phase with respect
Consider the all-pass lattice topology shown in Fig. to a steady-state sinusoid at its input. This critical fre-
7. It has the all-pass transfer function quency becomes the normalized center frequency r =
2~rfcT (with T being the sample period) for all filter
A(z) - Y(z) 13 -t- ~(1 q- 13)z-1 Jr- z -2 types employing the topology shown in Fig. 9.
X(z~) - 1 + ~/(1 + 13)z-1 + [~z -2" (6) In Fig. 9 we have introduced a new control coefficient
k. When k = 0, the network in Fig. 9 implements the
Some characteristics of the all pass filter are summarized notch [Eq. (9)] exactly, and when k = 2, this same
in Fig. 8 and the equations network implements the perfect resonator [Eq. (8)] ex-
actly. Within these bounds, this control (for k < 1) gives
[A(eJ'~ = 1, A ( - 1) = ! , A(e j~ = - 1 . us the ability to specify the depth of the cut, leaving the
(7) magnitude at the extremal frequencies equal to 1. Using
the same network for the resonator, we can control the
It is interesting that the non-minimum-phase all-pass depth of the skirts (when k > 1), leaving the absolute
filter will shortly become integral to a parametric filter
that is indeed a minimum-phase design. We also note
in passing that the transfer to Dr(z ) from the input com-
prises only the denominator (the poles) of A(z),
~ "Sk[A(eJ~)
Dr(g ) = X(z)
1 + ~/(1 + 13)z-1 q- 13Z - 2 . (oc=l
0)
-3 -2 -1 -2~''"~ ....
IHr(eJC~
1
Absolute
0.8

0.6 Fig. 8. All-pass radian phase responses.

0.4

0.2
1/2 '~~-, 1+-~1-kl
0.5 1 1.5 2 2.5 3 O)

X,z,
Fig. 6. Perfect resonator; Q = 2, coc = 1. Fig. 9. Cut, notch, or resonator type filter.

Y(z)

Fig. "7. Lattice second-order all-pass filter.

J. Audio Eng. Soc., Vol. 45, No 9, 1997 September 669

DATTORRO PAPERS

peak frequency magnitude at precisely 1. These actions These depth equations are easily deduced from Eqs. (10)
explain the unusual looking normalization coefficient at and (7) and are independent of the center frequency.
the output. The center frequency is unequivocally determined by
The action of the control coefficient k is characterized Eq. (10) for these cut, notch, resonator, and perfect
in Table 3 and illustrated in the magnitude-squared re- resonator filters shown in Figs. 10 and 1 1. This center
sponses of Figs. 10 and 11. The cut depth [Eq. (11)] frequency corresponds to the peak or trough extremum
and the skirt depth [Eq. (12)] must be squared to resolve of the magnitude transfer evaluated on the unit circle in
with Figs. 10 and 1 1, respectively,

1 - (1 - k) _ k
absolute cut depth = ;0~<k< 1 (11)
l+ll-kl 2-k

1 +(1 -k) 2-k

resonator absolute skirtdepth - 1 + I1 - k I - k ;1<k~<2 (12)

Table 3. Control coefficient k for Fig. 9.

the z plane.
k = 0 Notch, H,(z)
0 < k < 1 Cut, He(z) The ordinate axis is drawn at the lower half-power
k = 1 Yields Input signal excursion frequency oh for the plots in Figs. 10 and 1 1.
1< k < 2 Resonator, H b (z) The half-power excursion frequencies (the two musical
k = 2 Perfect r e s o ~ o r , H~(z)
cutoff frequencies) are given for the cut, notch, resona-

IHc(eJ~

O" 1

0.4

0 0.5 1 1.5 2 2.5 3 (.0

03c= 1
Fig. 10. Cut filter magnitude-squared responses for various values of k; Q = 2.

IHbnorm(eJ~ 12

1[A ~ k=6/5
k=7/5
o.st/i k=81s
k=2

0 6

0 0.5 1 1.5 2 2.5 3 ' ' (/)

Fig. l l. Resonator magnitude-squared responses for various values of k; Q = 2.

670 J. Audio Eng. Soc., Vol. 45, No 9, 1997 September

PAPERS EFFECT DESIGN

tor, and perfect resonators by the equations

(1 + [3)2 cos (t%) + , - ([3 - 1)N/2(1 + [32) _ (1 + [~)2 cos2((Oe)

cos (tOE,0 = 2(1 + 13z) (13)

[3 = 1 - tan (toc/(2Q)) = 1 - tan(Ato/2) and via a scaling by the boost factor k on the output,
(14) but only when k > 1. The transfer function of the circuit
1 + tan (toJ(2Q)) 1 + tan(Ato/2) "
in Fig. 12(b) is identical to that in Fig. 12(a). By pushing
the output coefficient forward, we simplify the other
Given a particular center frequency, the all-pass lat- coefficient.
tice coefficient [3 [Eq. (14)] precisely controls selectivity The action of the control coefficient k is now charac-
(the filter Q [Eq. (1)]) for these cut, notch, resonator, terized in Table 4 and illustrated in the magnitude-
and perfect resonator filters. 18 Whereas the lattice coef- squared responses of Fig. 13. The cut depth [Eq. (16)]
ficient ~/ is a function only of toc as we see from an and the boost [Eq. (17)] must each be squared to resolve
inspection of Eq. (10), here we see that [3 is a function with Fig. 13,
of both toc and Q as per our new definition of musical
cutoff frequency, Eqs. (3) and (2). Regalia absolute cut depth = k ;0~<k< 1 (16)
On the one hand, it is very good that we have discov-
ered closed-form mathematical relationships describing Regalia absolute boost = k ; 1 < k < oo. (17)
how to modify the two lattice coefficients to control the
musical filter parameters. But from a control standpoint,
we would like to have a way to decouple the filter coef-
ficients so that only one of them governs the center
frequency whereas the other governs only the selectivity
parameter. (We almost have that in ~.) Later on we will
see the Chamberlin filter topology, which nearly reaches l+k
that ideal. (a)
2.4.1 Regalia k Coefficients
In [13] it was understood that a simple algebraic
change in variable would result in a new design which 112
substitutes the parametric boost filter for the resonator,
hence incurring the loss of the resonator and the perfect
resonator. Employing the same topology as before, the (b)
coefficients in Fig. 12(a) are derived from those in Fig. Fig. 12. Cut, notch, or boost type filter. Transfer functions in
9 via the substitution (a) and (b) are identical.

1-k Table 4. Control coefficient k for Figs. 12 and 14.

1 - k---~- (15)
l+k
k = 0 Notch, Hn(z)
0 < k < 1 Yields Cut, He(z)
k = 1 Input signal
is This equation for 13 is exact in terms of the selectivity 1< k < oo Boost, Hb(Z)
definition, Eq. (1).

iHb(ejO))12 2
1.75

1.5

IHc(eJ('~ 2

0.5

0.25

; .... o'.5 " 1 ' 1'.5 .... ~ .... 2'.5' 3

(0
O)c= 1
Fig. 13. Cut and boost responses for various values of k; Q = 2.

J. Audio Eng. Soc., Vol. 45, No. 9, 1997 September 671

DATTORRO PAPERS

These results can be derived by subtituting Eq. (15)

into Eqs. (11) and (12). As before, these results are
2.4.3 Lattice Topology in Practice
independent of the center frequency. The combined plot The foregoing filter topologies constructed from the
in Fig. 13 highlights the symmetry of the cut with the all-pass lattice suffer two drawbacks to their implemen-
boost filters, hence, the symmetry of Q. tation: 1) the lattice produces spontaneous low-level au-
In general, the parametric filters of Figs. 9 and 12 are dible zero-input limit-cycle tones, and 2) the lattice to-
minimum phase. From the all-pass characteristics [Eq. pology is prone to signal overflow at internal nodes
(7)] it can be deduced that, independent of center before the all-pass output has reached full scale. 21
frequency, The first problem is solved by magnitude truncation
of 22 all lattice memory elements [9]. The second prob-
Regalia absolute skirt depth of boost filter = 1 lem is solved by scaling the lattice input, as already
shown in Figs. 9, 12, and 14. When premature internal
;1 < k < o o . (18) overflow persists (which is more likely for high Q, cut or
boost), it becomes necessary to provide a user-controlled
The ordinate axis is again drawn at the lower half-power input-signal-level adjustment. 23 Compensation will be
excursion frequency (the lower musical cutoff frequency required at the filter output, under separate user control.
too in Fig. 13. The two musical cutoff frequencies (o2,1 Keep in mind that the cost of output amplification is the
for the boost filter are derived using a slightly different concomitant amplification of the filter's internal signal-
Eq. (19) as compared to that for the resonator [Eq. (3)]. truncation noise floor. So this input scaling process
But it can be shown that the results are the same as should be limited.
before, that is, Eq. (13) remains valid under As an alternative to the use of the all-pass lattice, we
recall that the direct form I filter topology does not suffer
IHb(eJ=)12 - 1 1 from internal signal overflow. That is because its single
[Hb(eJ=c)[ 2 -- 1 = 2 " (19) accumulator has infinite headroom when used in a fixed-
point implementation [12, pp. 857, 875] 24, [11, ch.
Unlike Eq. (3), Eq. (19) does not reduce to the tradi- 11.3], [14, ch. 6.7.1]. In Fig. 15 we show an implemen-
tional definition of cutoff frequency because Hb(e joo) is, tation of the second-order all-pass filter [Eq. (6)] com-
by definition, never zero. Because we were able to derive prising embedded direct form I first-order all-pass sec-
tions. 25 This topology retains the dichotomy of the
the Regalia k coefficients in Fig. 12 from the network
in Fig. 9 while retaining the same topology, all the musical filter coefficients as in the lattice, while em-
equations thus far remain applicable, that is, for [3, % ploying the exact same coefficients, and avoids inter-
nal overflow.
toe, and 0 ) 2 , 1 .
Empirically we observe that the limit-cycle tones pro-
2.4.2 R e g a l i a - M i t r a T o p o l o g y duced by the direct form I are much quieter than those
Fig. 14 shows the parametric filter topology originally produced by the corresponding lattice in Fig. 7, in gen-
presented ~9 in [13]. 20 Although our development led us eral. 26 Truncation error feedback (not shown) in the di-
to Fig. 12, the transfer function of the circuit in Fig. 14
is identical. 21 Conditional saturation is helpful but does not solve the
problem because internal clipping sounds bad. In [ 11, ch. 4.3]
a novel topology for the perfect resonator is shown.
19 We moved to the input the originally internal scaling 22 See Part 3, Section 9.2, Appendix 8.
by t/2 in order to avoid subsequent overflow in a fixed-point 23 This knob is probably required anyway to counterbalance
implementation. boosts at the filter output.
2o Regalia and Mitra [13] also formulate the construction of 24 Overflow is not always a bad thing.
first-order shelving filters using the same topology. In the audio 25 Conversion to direct form II using Rossum's technique
industry, shelving filters are typically first-order designs. They [17] would eliminate some memory elements while providing
are used to uniformly boost or cut selected portions of the automatic input scaling. The scaling is necessary to prevent
high- or low-frequency region. They are called "shelves" be- internal overflow in that topology.
cause their magnitude responses approach unity asymptot- 26 The direct form I may require error feedback [12] to be
ically. truncation-noise competitive with the lattice, however.

1/2

Fig. 14. Regalia-Mitra topology.

672 J Audio Eng. Soc., VoL 45, No. 9, 1997 September

PAPERS EFFECT DESIGN

rect form is known to further minimize limit-cycle oscil- Eq. (27)], to the perfect resonator [Eq. (5)],2s
lation [18], thus providing an alternative to magnitude
truncation as a remedy. For both the lattice and the em- '/2(1 - [3)(1 - z -2)
bedded direct form, stability is assured by < 1 and HF(Z) =
1 + ~/(1 + 13)z- l + 13z-2
< 1.
q2(1 - fl)(l - z -2)
2.5 A p p e n d i x 2: Filter Errata in t h e Literature
1 - 2R cos (0)z- 1 + R 2 z - 2 9 (5)
A mistake has been perpetuated regarding the center
frequency of the second-order digital filter. The polar
Using Eq. (10), we can easily deduce the following
representation of complex conjugate filter poles is often
identifications;
found, correctly written, as

•=R 2
Zpole = Re -+j0 . (20)

The erroneous hypothesis can be recognized wherever - 2R cos (0)

~/- 1 +R 2 = -cos(t%).
the filter's normalized center radian frequency (o~ is as-
cribed to the radian pole angle 0. Hence, the distinction
between c e n t e r frequency and pole frequency is ob- This proves that the only instance where the center fre-
scured in the literature. 27 There it is argued that for high quency ~oc would be the same as the second-order pole
selectivity, this distinction is of little practical impor- (or resonant) frequency 0 is for conjugate poles right on
tance, but that tenuous assumption of practical equiva- the unit circle (R = 1). But in that circumstance one
lence has, consequently, promulgated specious theoreti- has an oscillator, not a filter.
cal conclusions within the audio community. These results can be extended to the resonator in gen-
One such erroneous conclusion is that the perfect reso- eral. Similar conclusions can be drawn from an examina-
nator transfer function [Eq. (5)], for arbitrary center tion of the second-order all-pole transfer function [see
frequency, does n o t h a v e a peak magnitude exactly equal Eq. (23)], and from the second-order all-zero transfer
to 1 when evaluated on the unit circle in the z plane. function, such as the one in Fig. 4.
The errant proof evaluates Eq. (5) at the resonant fre- An instance where the center frequency is identical
quency (at z = e J~ in complete disregard of the pole to the pole frequency is for the case of the first-order
radius R. Evaluation at the true center frequency (at z = resonator. The equivalence is independent of pole radius
e j=o) given by Eq. (10) shows that conclusion to be false, R, unlike the second-order case. This instance may be
that is, it is true that the perfect resonator as given by the reason for the propagation of the erratum regarding
Eq. (5) always has a peak gain of exactly unity. the second-order case. The transfer function of the first-
We can establish a correspondence between pole and order resonator is
center frequencies by equating the general denominator
of a second-order transfer function, written in terms of 1-R
the pole radius and angle [Eq. (20)] [11, ch. 4.3], [12, Fr(z) - 1 - ReJOz- 1 9

This filter has only one pole. But notice that the one filter
27 The resonant frequency is that frequency at which a filter
rings when excited by an impulse. The resonant frequency is
the pole frequency, which is the same as the pole angle 0 in
the z plane. The center frequency is the frequency at peak 28 The poles occur in complex conjugate pairs when the filter
magnitude response in the steady state, when a filter is excited coefficients of z (that is, ~/ and 13) are real. When the filter
by a sinusoid of infinite duration. In general, center and reso- coefficients are real, then it is easily shown that the filter's
nant frequencies are not identical [19, ch. 5.5]. impulse response must also be real.

X(z) 13 Y(z)

Fig. 15. Embedded direct form I, second-order all-pass filter.

J AudioEng. Soc., Vol. 45, No. 9, 1997 September 673

DATTORRO PAPERS

coefficient is complex, in general. Hence the impulse zeros serve to provide high attenuation there. In contrast,
response of this filter cannot be real. One may surmise musicians have a taste for peaked filters, even when the
that there must be some interaction among multiple poles desired filter is of the low-pass variety. Because the
in the z plane, which destroys radial symmetry. musician's peak-center frequency is typically quite low
(requiring poles closer to the unit circle), zeros are
3 CHAMBERLIN FILTER TOPOLOGY largely unnecessary due to the relatively high attenuation
at frequencies far away from the low-frequency poles.
Next we consider high-fidelity musical filtering using When the peak-center frequency is high, on the other
a different topology and the musician's all-pole low- hand, the stopband excursion of the all-pole filter magni-
pass filter type. We apply our earlier redefinition of tude response may not span 3 dB. In fact, when the
cutoff frequency, in terms of half-power excursion, to peak-center frequency reaches "rr/2, the all-pole low-
this new construction, which establishes a tie to our pass filter ceases being low pass because the magnitude
previous work. response at "tr starts to exceed the response at dc.
The musician's all-pole filter has antecedents in the Due to the fact that the Chamberlin filter is all pole,
electronic music industry, 29 appearing in currently re- there is little control over the rate of transition from
nowned and vintage music synthesizers [22]. The filters passband to stopband. To increase the transition rate of
we considered previously had zeros in the transfer func- the low-pass filter, the accepted solution is to cascade
tion. We were concerned about the control of those filters an identical all-pole filter. This works in practice be-
as a musician might like to control them. Here we present cause the musician's working range of the low-pass
an additional goal; namely, to come up with filter coef- peak-center frequency is much less than r for reason-
ficients where each will control individually only center able sample rates. Zeros placed at the Nyquist fre-
frequency or selectivity (filter Q). To do so, we rederive quency, for example, would have little impact over the
the Chamberlin [23] all-pole (two-pole) low-pass filter musician's working range. Therefore the cascade is pre-
topology entirely from the perspective of the discrete- ferred to zeros at Nyquist. Zeros elsewhere in the stop-
time domain. 30 band region would entail more computation, hence they
We argue that the truncation noise performance of are undesirable. In this development, we will consider
the Chamberlin filter topology is very good, although only a single filter section.
practitioners have known that for years. 31 In so doing We expect some kind of boosting response, as shown
we introduce a new more musical and conservative mea- in Fig. 16. The corresponding transfer function must be
sure of noise performance that we call "transparency," at least second-order to get the peak center away from
and which we denote criterion 1. Using a more tradi- dc. Notice that the filter is normalized to unity at d c . 33
tional approach, denoted criterion 2, we compare the Once again, we must refine our notion of cutoff fre-
truncation noise power observed at the Chamberlin filter quency by relating it to half-power excursion, as before.
output to the input-signal quantization noise power, For this filter type, we define the passband excursion
which can be construed as the noise gain. We discover from the value of the magnitude-squared response at dc
that for the Chamberlin topology, the worst noise gain to the peak value of the response. Reminding ourselves
is the same as the peak gain squared of the whole filter that this magnitude-squared response is periodic in 2~r,
acting upon the input signal. That turns out to be the we then similarly define the stopband excursion from
reason why the noise performance is so good. the peak to the value at Nyquist. 34 In Fig. 16 the half-
power excursion points are indicated, defining the musi-
3.1 S h a p e of t h e M u s i c i a n ' s L o w - P a s s Filter cian's bandwidth of the all-pole low-pass filter.
The electronics engineer's low pass has zeros in the We find the frequencies of the half-power excursion
stopband and is very fiat in the passband. 32 The stopband points (the musical cutoff frequencies) here much like
we did before: the passband half-power excursion fre-
quency is found solving Eq. (21) for to,
29 The classic Moog analog synthesizers, for example, em-
ployed fourth-order all-pole voltage-controlled filters (VCFs). IHchx(eJ=)l 2 - 1 1
His constant-Q design was also known as the Moog ladder, IHchx(eJ=o)12 -- 1 = 2 " (21)
after the appearance of the schematic [20]. A cascade of two
Chamberlin filters can be considered as the digital counterpart
to the Moog VCF because some of the same characteristics
are shared. They are both all-pole constant-Q designs tuned by We call this frequency to1- Similarly, we call to2, the
a single sweepable parameter. Rossum [21] of Emu considers solution to Eq. (22) for the half-power excursion in
essential nonlinear ingredients to make digital filters sound
more "analog."
3o This filter was originally derived from an analog state-
variable filter by application of the impulse-invariant trans- 33 To bring the boost at the peak-center frequency toe down
formation.. to unity, additional scaling is required beyond what we recom-
31 The Chamberlin all-pole design is a reputed resident mend here.
within the contemporary digital synthesizers by Peavey and 34 The electronics engineer's transition band and stopband
Kurzweil. are merged in this development. Because of the lack of zeros
32 The Butterworth filter, for example (which is a good here, the electronics engineer's boundaries are not as clear.
choice for audio with regard to minima/ringing), has all its Also, the electronics engineer would measure bandwidth from
z e r o s at Nyquist. de, unlike our measurement.

674 J. Audio Eng. Soc., Vol. 45, No. 9, 1997 September

PAPERS EFFECTDESIGN

the stopband, At the peak-center frequency the magnitude-squared

response reaches its peak height. Exactly,
[H~hx(eJ=)12 -- IH~h~(-- 1)12 = _I
(22)
[Hchx(e-i'o)[2 -- IHchx(- 1)12 2" 40t213
m~x[iHchx(eJ~O)12
] I- I = (13 _ 1)2(413 _ k z)
Neither of these two cutoff frequency definitions [Eqs.
(21) and (22)] reduce to the traditional definition because a2(1 + 13)2
none o f the terms can go to zero in this all-pole design. = (13 - 1)211 + 132 -- 213 cos (2to~)] "
(25)
3.2 Transfer Function Development
We begin with a simpler second-order transfer func- For the low-pass filter we normalize the transfer function
tion having no zeros, so we can expect some of the to unity at de, so ct becomes
previously discovered equations to be different,
ct = 1 + h + 13.
at
Hehx(Z) = 1 + hz -1 + 13z - 2 " (23)
The two musical cutoff frequencies were determined
exactly, using M a t h e m a t i c a [24],35 as

cos 2, sin 2 (o~d2)(13 - 1) %/211 + 132 _ 213 cos (2c%)]

COS ((02,1) = COS ((t)c) -[-, (26)
%/1 + 13 + 8[32 + 133 + 134 -I--,-- 413(1 + 13)ZCOS (tOe) -- 13(1 -- 6[3 + 132)COS(20~c)

We were not able to determine an exact expression for

We seek the relationship of the ideal coefficients h and the 13 coefficient in terms of toe and Q as we did for ~/,
13 to the peak-center frequency o~ and selectivity Q, but the following guess turns out to be a good
approximation:
[-(1 + 13)h] (24)
toe = arccos 413 " 1 - sin (toJ(2Q))
[3 ~ 1 + sin (~%/(2Q)) " (27)
If we express h as
The plot in Fig. 17 shows that our expression for 13
is good over the r e c o m m e n d e d operating peak-center
k- 413~/
1+13 frequency range of toc -- 0 to "rr/2. To m a k e this plot,
we substitute the desired Q into the exact equations for
then we find the simpler expression for peak-center to2 and (o 1 [Eq. (26)] using the approximation to 13 [Eq.
frequency, (27)], and then we sweep over toc. If we had the exact
expression for 13, then the sheet would be a taut plane
~ = arccos ( - ' y ) . (10) having unit slope with respect to the desired Q. But the

This equation for the center frequency is the same as 35 The extensions of Mathematica [25] to analog and digital
before, and both Eqs. (10) and (24) are exact. signal processing are highly recommended.

IHchx(CJ00)12
2 peak2 = 1.961444

1 . 5 (0631012,1 ~ 2 2 ) ~

defining bandwidth
k
O. 5

O-~e= 1
skirt depth2 -- 0.166305
0.5 1 1.5 2 2.5 3 (0
Fig. 16. All-pole low-pass magnitude-squared response.

J. Audio Eng. Sac., Vol. 45, No 9, 1997September 675

DATTORRO PAPERS

approximation Eq. (27) is far better than some others in the z -2 coefficient in Eq. (23) are
the literature [ 2 6 ] - [ 2 8 ] , [15, ch. 2-111, p. 123]. The
largest percentage errors in the recommended center fre- 1 1 5 3+ 1 4
quency range are for a desired Q of 1, having error 13 1 - + - 24Q3 0% 1 2 ~ 0%
- -
~d z~d-
maxima of 21.8 to 27.6%.
61 5+
3.2.1 Approximations 1920Q5 o~c 999.
To achieve our stated goal of obtaining filter coeffi-
cients that control center frequency or filter selectivity These series are hard to predict. The Mathematica script
individually, we now make series approximations to our used to generate them is

beta = (1 - Sin [wc/(2 Q)]) / (1 + Sin [wc/(2 Q)]) ;

Simplify [Series [Simplify [Factor [ - 4 beta Cos [wc]/(1 + beta)]], {wc, O, 5}]]
Series [beta, {wc, O, 5}]

Fig. 18 shows the musical filter topology36 that imple-

expressions for the ideal filter coefficients we have found ments a truncated series approximation to the ideal filter
thus far. Using the good approximation Eq. (27), we coefficients, hence decoupling somewhat the control of
find that the first several terms of the equivalent Maclau- c% and Q. Fig. 18 incorporates the first three terms from
rin series for the z-1 coefficient in Eq. (23) are the h series and the first two terms from the 13 series.
Thus the coefficient F~ is identified with co~ while the
coefficient Qc is identified with 1/Q. Because the circuit
~ 0c -- ~0c + O)c implements so few terms from the 13 and X series which

+~ ~
4(1 2-~ + + 19 (23 ~% "'"
) 36 We adopt Chamberlin's nomenclaure [23]. Chamberlin
points out that this filter topology simultaneously possesses a
high-pass and a band-pass output at the nodes labeled hp and
bp, respectively. We discuss only the low-pass filter function
and the first several terms of the Maclaurin series for lp of this circuit in detail here.

OY,,

lq

desired

.5
mc
Fig. 17. Actual all-pole filter Q as a function of center frequency and desired Q.

X(z)---~ hp F~ b Fc
Y(z)

i
Fig. 18. Chamberlin topology, second-order all-pole filter. Input scaling by 1/2 and output compensation not shown (see Sec-
tion 3.3.6).

676 J Audio Eng Soc., Vol. 45, No. 9, 1997 September

PAPERS EFFECT DESIGN

are themselves approximations due to Eq. (27), these peak-center frequency o~ [Eq. (24)] in this range, we
stated identities are crude. We refine the approximate discover in Eq. (31) that Fr and Qc are not completely
relation of Fc to o~r in Eq. (29), but we will leave the decoupled, 39 except for very high selectivity (Q~ ~ 0),
circuit in Fig. 18 as it is. After we characterize the circuit
a little more, we will find that the filter coefficients in 0 < cos (toe) =
the figure provide sufficiently autonomous control for
musical purposes. 3
4(1 -- F~Q c) - F~(2 - a~) + F~Q~ ~< 1. (31)
The all-pole low-pass transfer function of this further 4(1 - F~Q~)
approximation to Eq. (23) in Fig. 18 is

Y(z) F~z- t
nch(Z) -- z - 1H~hx(Z) (28)
X(z) 1 - (2 - FcQ ~ - F~)z -l + (1 - FcQc)z -2

redefining
k = - ( 2 - FcQ~ - F 2 ) , [3 = 1 - FcQ c (29) The identity in Eq. (31) is exact. Further, we find on
the right-hand side inequality that
where
2
Fc<~c-Q~
Fe ~- 2 sin rad, Qc = ~ .
and on the left-hand side,
The transmogrified numerator of Eq. (23) now shows a
delay operator in Eq. (28). This comes about because - Q ~ + %/8 + Q2
of the need to eliminate an otherwise delay-free loop in Fc <
the circuit of Fig. 18. 37 The numerator coefficient o~ has
become F 2 to force Eq. (28) to unity at dc (z = 1), as From the stability condition Eq. (30), the minimum
stipulated. This refined approximation to F~ in Eq. (29) value of Fc is zero. This is achieved for the right-hand
is from [23], [26]; it becomes more exact for high Q. side inequality of Eq. (31) when Qr reaches ~v/2. Thus
When the peak-center frequency t% is low and Q is high, we have an upper bound on Qc to keep the actual peak-
Fr in Eq. (29) reduces to t%. [An exact expression for center frequency within the prescribed tuning range,
Fr is given by Eq. (31).]
0 < Qr < ~v/2.
3,2.2 Stability~ParameterDecoupling
The stability of complex conjugate poles demands To maintain complex conjugate poles in Eq. (28), 40
the constraint we find that the following inequality holds:

0~< (1 - FeQ~) < 1 . 0~ (F~ + Qc) ~< 2 .

Restated, This is found by rooting the denominator of Eq. (28),

0<F~ <- 1 . (30) F2z- 1

a~ Hob(Z) = (1 -- az-l)(1 -- a*z -1) (32)
This condition is ascertained from Eq. (28) by de-
manding a pole radius 38 of magnitude less than 1. where
From previous considerations we presume that the tun-
ing range for the all-pole low-pass filter is a = 1 Fc(F~ + Q~) + jFr (F~ + o'~..r
2 u 4
,IT
~>~o~>0.
Using the upper bound we found for Qc, we see that
there will always be some finite range of F c over which
I f we substitute Eq. (29) into the equation for the actual
the poles will be complex conjugate.

37 It is remarkable that the delay-free loop is eliminated

without compromise to the digital filter coefficients, because 39 A similar conclusion can be reached by solving Eq. (27)
delay-free loops can be troublesome when it is desired to main- for Q in terms of Fc and Qr via Eq. (29); we leave that for the
tain autonomy of the coefficients in an analog-to-digital fil- reader. But note that Eq. (27) is an approximation whereas
ter transformation. Eq. (31) is exact.
3s See Eq. (5) in Section 2.5, Appendix 2 to find the pole ~0 That is, for peak-center frequency away from but asymp-
radius. totically including de.

J. Audio Eng. Soc., Vol. 45, No. 9, 1997 September 677

DATTORRO PAPERS

Combining all three criteria, we finally conclude that Q filter. Instead we will repeat the musical analysis, as
to maintain a stable low-pass filter in the form of Eq. in Fig. 17, relating Q and center frequency; but this time
(28), having complex conjugate poles conforming to the we will not use the ideal coefficients. Rather we use the
prescribed tuning range, then the constraint must hold: actual filter coefficients given by the truncated series
approximations in Eq. (29).
2 Specifically, the radian frequencies <oc, to2, and o~1 in
0 < F~ < rain 1,2-Q~,~-Qr Fig. 19 are calculated using the actual filter coefficients,
evaluating Eq. (31) to get o~c, and by substituting the
03) truncated series approximation for 13 [Eq. (29)] into Eq.
(26). Fig. 19 tells the whole story by relating actual
We learn from Eq. (33) that an artificial upper bound filter Q [Eq. (1)] to the filter coefficients. Ideally, we
on the value of Qr equal to 1 yields a universal upper are looking for a planar relationship. Nonetheless, the
bound on Fr equal to 1 as well (corresponding to toe of sheet is fairly unwrinkled up to Fc ~ 1, corresponding
about ,rr/3). We conclude that we can guarantee stability to a tuning range of peak-center frequencies up to "tr/3.
of complex conjugate poles for any value of either filter Further, it appears that for our purposes the selectivity
coefficient as long as they individually remain within parameter control Qc is sufficiently decoupled from the
the range of 0 ----> 1. tuning frequency control F c. Hence we can expect good
agreement between theory and practice in that region. 41
3.2.3 Peak Gain
We examined the actual peak gain of H~h(ej'~ (evalu- 3.3.1 Integrator Analysis
ated at o~) over the prescribed ranges of Fc and Q~ (both Generally speaking, it is not a good idea to implement
0 --> 1) substituting the truncated series approximation an ideal digital integrator unless it can be guaranteed
coefficients [Eq. (29)] into Eq. (25), then taking the that there exists a zero of transmission across it at dc.
square root. We found the peak gain to be greater than This is certainly the case for integrator 1 in Fig. 18,
but approximately equal to 1/Qr The largest excess be- which has the required zero across it, but integrator 2
yond this estimate occurs for low center frequency and has no such zero. In that case one must then prove that
low Q, or for high center frequency and high Q. At there can exist no signal from any source having dc
Fr = 0.000001 and Q~ = 1 we found the greatest excess content upon arrival at the input to the integrator under
at about 15.5% more than 1/Q~. scrutiny. Audio signals normally enter the digital circuit
There is no separate control over peak gain in the at the designated input node, but noise having dc content
Chamberlin topology; it is controlled indirectly through is routinely generated in any practical implementation at
Qc. We recommend a maximum peak gain of 24 dB for every node where signal truncation occurs. These noise
musical purposes, corresponding to a minimum Qc of sources often appear in contemporary DSP chips at the
about 0.0625 (filter Q = 16). input to each multiplier because multiplier inputs cannot
accommodate double-precision operands the way the ac-
3.3 Performance of the Chamberlin Filter
Now we wish to know whether our approximations
are good enough. To do this, an engineer might calculate 41 At a sample rate of 44.1 kHz, w/3 corresponds to a band-
width of 7350 Hz. Considering that the topmost note of the
the root locus of the Chamberlin poles to see how closely pianoforte reaches only 4186 Hz, that tuning range is good
their trajectory matches that of a second-order constant- enough for musical purposes.

1/Q c

2
Fc
Fig. 19. Actual Chamberlin filter Q as a function of F c and 1/Qc.

678 J. Audio Eng Soc., Vol. 45, No. 9, 1997 September

PAPERS EFFECT DESIGN

cumulators can. 42 The high-rate noise is accurately mod- topology it undergoes filtering as does the signal itself.
eled as a deterministic source [12, Eq. (6)], input to a Using the truncation noise model described for the integ-
fictitious adder resting in front of the multiplier. Fig. 20 rator analysis, we then need to know the transfer func-
demonstrates the application of the noise model to one tion from each of the internal noise sources to the low-
of the noise sources (e2) on its way to integrator 2. pass output. Having obtained this information, we can
For the Chamberlin topology we have the remarkable predict the frequency-dependent amplification of each
result that the input to integrator 2 never sees any signal presumably wide-bandwidth noise source.
having dc content. For verification, we now look at the We assume that all internal truncation occurs at the
most interesting signal, which is the noise source at the same bit level. Suppose, for example, that internal noise
input to the multiplier at node bp. There we have due to truncation were being generated at the 20-bit
spectral level. Then any given noise magnitude response
12(z) _ F~ 1 - (2 - FcQ~)z -1 + (1 - F~Qc)z -z would need to possess at least a 24-dB boost beyond
e2(z) A unity (in any frequency region) before criterion 1 were
(34) violated, assuming 16-bit input signal fidelity and a
spectrally fiat signal-quantization noise floor. 43
where A is the denominator of Eq. (28). The transfer Like the one shown in Fig. 20, there are three noise
function (34) has a zero of transmission at dc, which sources arising due to truncation at each of the respective
can be proven by substituting z = 1. The three other multiplier inputs. These nodes are labeled hp, bpq, and
possible sources (at nodes hp, bpq, and the filter input) bp in Fig. 18.
acquire a simple zero of transfer at dc in the form 1) hp: The noise transfer function from hp to the
1 - z - 1 by the time they arrive at 12. output is --Hch(Z ). This means that the noise transfer
from hp is the same as the signal transfer (with sign
3.3.2 Truncation Noise; Spectral Criterion inversion), which is certainly not a bad situation. This
The object of our noise analysis is to find the internal source is the largest of the noise contributors when ob-
truncation noise generated by the circuit itself, which served from the filter's low-pass output.
then appears at the observed output. The design goal of 2) bpq: The noise transfer from bpq to the output is
fidelity might be stated as follows: QcHeh(Z). We know that the peak gain of Hch(ejc~ is
approximately 1/Q~, which means that the peak of this
Criterion 1: It is desired that the filter circuit generate noise magnitude response from bpq is about 1. This
no truncation noise appearing at the filter output which is good.
would fall above the spectral noise floor due to quantiza- 3) bp: Lastly, we consider the noise transfer from
tion of the original input signal. bp to the output, which can be written

Under this criterion it would be a violation for any 1 - z -1 Q~) H~h(Z)

portion of the truncation noise spectrum created by the _ Fcz- I
filter itself to fall above the presumed spectrally white
quantization noise floor of the input signal. Hence this When filter Q is high, the first term introduces a zero at
criterion is the more conservative of the two criteria for dc (but only to the noise), which is exactly what happens
fidelity that we will consider. Under this musical and when we employ first-order truncation error feedback
spectral interpretation of fidelity, the filter is then termed [12]. The zero squelches the noise in the low-frequency
transparent to the input signal. region, but boosts it in the high-frequency region. Using
Because the noise we are studying is deterministic Eq. (29) we find that the crossover point (the frequency
(the sources are known), when it occurs early in a filter at which I1 - z-ll/Fc = 1), above which the 1 - z -1
term begins to boost, is approximately r In this partic-
ular circumstance, the second-order low-pass filter
42 We presume no truncation post-accumulation for these H~h(Z) kicks in above 00c to remove the boosted noise.
integrators. This presumption is justified based on the alterna- Hence the center frequency and peak gain of this noise
tive, which is a leaky integrator, requiring a multiply in its
loop. source are about the same as that of hp, for high Q.
Summarizing the three noise sources, the noise source
ez[n] at hp would demand 1 extra bit in the filter signal path
for every 6 dB of peak gain at the filter output (Section
-~. Fc 3.5, Appendix 3), because node hp sees the same gain
to

43 (20 -- 16 bits) • 6 dB/bit = 24 dB. Another way of

looking at this example would be to say that under fidelity
criterion 1, four extra bits beyond 16 would be required in the
filter signal path to maintain filter transparency if any one of
the noise magnitude responses were capable of a 24-dB gain
in any frequency region, that is, 1 bit for every 6 dB of gain
beyond unity magnitude response. Refer to Appendix 3 in
Fig. 20. Truncation noise source model (Appendix 8). Section 3.5 for supporting noise concepts.

J. Audio Eng. Soc., Vol. 45, No. 9, 1997 September 679

DATTORRO PAPERS

as does the input signal. The noise source at bpq is what is commonly termed the noise gain [29, ch. 9.2.2].
insignificant when compared with the other two. Node This is essentially an estimate of the total power boost
bp is much like node hp. We finally determine the re- of the internally generated, presumed spectrally white
quired total number of extra bits as an uncorrelated sum noise. Any boost beyond unity is bad news, whereas
of the respective peak gains; any gain of unity or less is good. 45 The calculation of
noise gain is often performed in the frequency domain
exploiting the Parseval energy relation [Eq. (35)], which
~-log2V'l + (162 + 12 + 162) = 4.5 bit.
integrates the magnitude-squared response of a noise
transfer function,
So for a recommended maximum peak gain of 24 dB
(16 absolute) across the filter, under criterion 1, five oo 1 1 2~r
more bits would be required in the single-precision sig- G= ~ Ih[n]la--~j0 In(eJ=)ladto.
n= -oo
(35)
nal path.
Hence the result of the calculation is unitless since the
3.3.3 Truncation Noise; Power Criterion integrand is a ratio. From the noise transfers we know
What we have thus far are the deterministic noise that the worst offender is due to the noise source at hp
transfer functions, which were of interest because of the in Fig. 18. Substituting the poles of that transfer [Eq.
way that the first criterion for fidelity was stated. The (32)] into the integration results from [14, ch. 4.7, p.
design goal of fidelity has a second, quantitative (statisti- 187; ch. 6.9.1, p. 357], we calculate the noise power
cal) interpretation which is stated as follows: gain at hp as

Criterion 2" It is desired that the truncation noise F4{a[1 - ( a * ) 2] - a * ( 1 - a 2 ) }

power generated by the filter circuit, observed at the Ghp :" (a - a*)(1 - a2)[1 - (a*)2](1 - aa*)"
filter output, be less than the noise power due to quanti- (36)
zation of the original input signal.
The vertical axis in Fig. 21 represents Ghp evaluated
Here we compare internally generated truncation noise over the recommended operating ranges of F c (0 ---> 1)
power observed at the filter output to the signal- and Qc (1 --~ 0.0625). There is zero noise gain for Fr =
quantization noise power observed at the input. This 0, because the filter is shut down at that point. We see
criterion is a classical high-rate interpretation, 44 ubiqui- that the worst noise power boost occurs at hp, for high
tously ascribed as the - 6 - d B average noise power per center frequency and high Q, where Ghp reaches
individual bit [14, ch. 3.7.3]. We begin by assuming 10.7826, which translates to 1.8 bit,
that the number of bits in the filter's single-precision
signal path is the same as the number of bits representing 10 log (1 + Ghp) = log2 V1- + Ghp
the input signal. We then determine the number of extra 20 log 2
bits required in the signal path to maintain input signal
fidelity using this criterion. To do so we must calculate = log2N/l + 10.7826 bit.

45 A noise gain of 1 says that the noise transfer function

44 The term "rate" here refers to the number of bits per will not increase the amount of noise that it passes, but says
sample. nothing about the spectral distribution of the noise passed.

Qc
0 0.
0~ ~

1 O.

% i0
7.

9 0.4
0.4 0 . 2 ~ - ~ ~
0
Fc
Fig. 21. Showing Ghp only where it exceeds unity.

680 J. Audio Eng. Soc, Vol. 45, No. 9, 1997 September

PAPERS EFFECT DESIGN

Gbp looks much like Ghp. The aggregate of the noise- truncation noise analysis of the Chamberlin topology
power gain contributions, corresponding to all the pre- we see that only one of the truncation noise transfers
sumably uncorrelated noise sources in the circuit of Fig. (bp) has such a term. Hence limit-cycle tones cannot
18, is calculated simply as the s u m Ghp d- Gbpq -]- Gbp be completely ruled out, although empirically we do
when all signal truncations occur at the same bit level. not recall overt problems with the circuit behavior in
This sum demands 2.2 bit, worst case as above: that regard.
We have relaxed the mathematical rigor in this section
log 2 X/1 + Ghp + Gbpq + Gbp because limit-cycle analyses for arbitrary topologies are
analytically difficult, in general. The question remains
= log2X/1 + 10.7826 + 0.0421196 + 10.4769bit. as to whether limit cycles are a serious problem for the
Chamberlin topology. Further analysis is certainly called
We conclude that the aggregate requires a total of three for. In the meantime we design for the worst case. So
extra bits (beyond the desired fidelity) in the filter signal our recourse is to minimize the potential tones' ampli-
path to maintain fidelity in accord with criterion 2. tudes of oscillation by providing internal signal trunca-
tion at lower bit levels. That is tantamount to providing
3.3.4 T r u n c a t i o n N o i s e S u m m a r y a higher precision signal path. A good rule of thumb is
From the perspective of either criterion, the all-pole that there exist about 6 dB of limit-cycle suppression
low-pass Chamberlin topology looks very good from the for each appended bit of precision.
standpoint of truncation noise performance. This is true
because the pole gain, which is the determinant of noise 3.3.6 Signal Overflow Analysis
gain in general, does not exceed the desired filter peak The study of overflow is concerned with the observa-
gain for the Chamberlin topology; the maximum pole tion of the signal magnitude at sensitive nodes. Typi-
gain is the maximum peak gain, which we recommend cally, one or several sensitive internal nodes may over-
to be 24 dB. flow (or underflow) sooner than the output. A saturation
nonlinearity clips (appropriately full-scale positive or
3.3.5 Limit-Cycle Oscillation negative) the overflowed node, as this is highly prefera-
Zero-input limit cycling arises due to ongoing signal ble to a two's complement wraparound nonlinearity. The
quantization within a recursive topology [ 11 ] - - a nonlin- audible consequence of clipping at internal nodes is
ear operation in an otherwise discrete linear system. 46 much more objectionable than clipping at the filter out-
The quantized filter coefficients are parameters to limit put, however, so it must be precluded completely. The
cycles, but are not the cause. Limit cycles manifest them- sensitive internal nodes are, once again, the multiplier
selves as annoying low-level tones at a circuit's outputs inputs because they typically cannot accept overflowed
after input signal has been removed. Signal quantization inputs like the accumulators can. 49 These are labeled hp
in most modern DSP chips often takes place at the single- and bp in Fig. 18.
precision multiplier inputs where double-precision op- In our overflow analysis, what we are really interested
erands cannot be accepted and so must be truncated. 47 in is the relationship of the sensitive nodes to the output.
The limit-cycle tones can therefore be visualized to enter So we form a ratio R of transfers to the sensitive nodes
the topology at the same places as the truncation noise. with respect to the transfer to the output node.
One such input port is shown in Fig. 20. Similarly to 1) Node hp: Formulate [Eqs. (28) and (29)]
the truncation noise sources, if limit-cycle tones occur
early in a filter topology they will be filtered just like hp(z)/X(z) _ (1 - z-1)2
Rhp(Z) --
H~h(z) FEz -l
the signal (at the same point of entry) itself.
In the decade just passed, we have learned that limit-
cycle oscillation is minimized by truncation error feed- _ sin 2 (o)/2)
back [ 18], which was devised to minimize the amplifica- [Rhp(eJ'~ sin 2 (o~J2) "
tion of truncation noise [ 12]. Essentially, error feedback
introduces zeros strategically placed on the unit circle This describes a boost over the output at high frequen-
into the noise transfer function, but leaves the signal cies. The worst case of overflow comes at the highest
transfer function alone. Therefore, a reasonable hypoth- frequency (z = - 1) and for a peak-center frequency r c
esis is that with or without error feedback, if the noise at the top of its utmost recommended range (~/2), for
transfer from a quantizer to the low-pass output has a
term 1 - z -1, then it provides some immunity to limit
cycles as well as some squelching of truncation noise,
both artifacts caused by that same quantizer. 48 From our 48 In fact, Laakso et al. [30] show that any zero in the noise
transfer function provides some limit-cycle immunity. The
common solution to both artifacts suggests that the two phe-
nomena are homologous.
Signal quantization converts a discrete-time system to a 49 Recall that most contemporary fixed-point accumulators
digital system. are designed to tolerate infinite intermediate output overflow
47 Multipliers then produce double-precision results, which simply by virtue of nonsaturating adders [11, ch. 11.3]. So
are usually fed to accumulators that can accept double-preci- saturation at an accumulator output (when necessary) is never
sion inputs. performed upon intermediate accumulated results.

J. Audio Eng. Soc, Vol. 45, No 9, 1997 September 681

DA'I-I'ORRO PAPERS

there the boost over the unity output is by the factor 2. requirement. Note that under the more conservative fi-
2) Node bp: Formulate delity criterion 1, the estimate of the required total num-
ber of bits in the filter signal path becomes 23 (n = 5,
bp(z)/X(z) (1 -- Z -1) r = 0).
Rbv(Z) -- Hch(;~) Fc The integrating accumulators must retain double pre-
cision feedback to maintain stability.
= sin (o0/2)
IRbp(eJ'~ sin (o0c/2) " 3.5 A p p e n d i x 3: T r u n c a t i o n N o i s e S p e c t r a l Level
versus Noise Power
Similarly, the worst-case boost over the output is abso- We seek the relationship of the truncation noise power
lute ~/2.
Based upon this analysis, a simple technique to elimi-
to
spectral level In(eJ=)12/M the noise power N because
we wish to prove that for every additional 6 dB of S/N,
nate internal signal overflow, which we have found the average noise spectral level drops by the same
works quite well, is to precede the Chamberlin topology amount. The analysis of truncation noise is much like
with a fixed input level attenuation of 1/z and to follow that of quantization noise [14, ch. 6.9.1, p. 353]. It
with a compensation factor of 2 at the output. The output is interesting that the classical high-rate estimation of
compensation amplifies the filter's internally generated quantization noise [14, ch. 3.7.3] is statistical in nature,
truncation noise, however, that is, the ratio of the input hence does not include the actual sample rate F s in its
signal power to the filter's own noise power degrades quantification. We therefore expect our final result to
by 6 dB. reflect this.
As shown in Fig. 16, the Chamberlin low-pass filter Here we regard truncation noise as deterministic so
is a boosting filter. For some input signals the low-pass that it has an integrable spectrum, and we note that a
output may overflow. But overflow at the output will discrete unity-level complex sinusoid of any duration
not always occur because the filtered signal may not and frequency has finite power S = 1. We then pose
have significant energy in the frequency region of the the problem: given 10 log (S/N) = 96 dB, 5~ input signal
boost. Further, some small amount of clipping at the duration equal to M samples, and input signal power S =
output is not offensive to the musician. Hence is it not 1, find the average truncation noise power spectral level
desirable to automatically normalize the filter peak gain "q21M in relation to the noise power N.
for the musician by attenuating the input signal because We easily find the noise power, solving
there will be an objectionable loss in perceived volume.
For then, the musician would demand a knob for output
10 log (N) = - 9 6
compensation. Such a knob is emphatically discouraged
because of the consequent amplification of internally
generated truncation noise. where
The most viable solution to the output overflow prob-
lem is to provide a filter input signal level user control. N= Td("~YIn(eMYr)12
e 0- - = 2-~
j 0 1(2= I'q(eJ=)12Mdo0
The user then determines at what input level any clipping
at the output becomes offensive. When an input level
user control exists for a boosting filter, it becomes un- (37)
necessary to provide user-controlled output compensa-
tion to maximize the output signal level. where ~o = 2~rfT throughout this paper. Eq. (37) 51 is
a statement of noise power versus noise power spectral
3.4 Estimate of Signal Path Width level, where T = 1/F s and M is the number of samples
The 21-bit path-width estimate given in Table 5 main-
tains input signal fidelity of 16 bits at the Chamberlin
low-pass filter output under criterion 2. The filter inter- 50 The approximate expected signal-to-noise ratio for a 16-
bit fidelity signal is 96 dB (6 dB per bit) [14, ch. 3.7.3].
nal signal path width can be minimized by reducing the 51 Scaling of the Parseval energy relation [Eq. (35)] by 1/M
maximum peak gain or by compromising the bit-fidelity to yield power in Eq. (37) is discussed in [31].

Table 5. Estimate of minimum required intemal signal path width at 24-dB maximum peak gain.

Bit Budget Attribute

N = 16 Output fidelity, assumed input signal quantization
n=3 Truncation noise immunity (criterion 2)
o=1 Internal signal overflow prevention
m=l Limit-cycle suppression (6 dB per bit)
r = max[0, (24/6) - n - m] Signal path LSBs for user-controlled input level attenuation*
N + n + o + m + r = 21 bits Total
* Assuming that the filter internal signal path resolution ultimately exceeds the input signal resolution,
then no input signal information will be lost through the use of an input level control, provided that
the attenuation is limited to the difference in resolution (6 dB per bit, 24 dB recommended).

682 J. Audio Eng. Soc., Vol 45, No. 9, 1997 September

PAPERS EFFECT DESIGN

in the original data record. If we designate ~2/M to be Signal Processing (Prentice-Hall, Englewood Cliffs,
the average noise power spectral level (the average of NJ, 1978).
]'q(eJ=)12/M over frequency), then solving Eq. (37) we [11] L. B. Jackson, Digital Filters and Signal Pro-
find that cessing, 3rd ed. (Kluwer Academic, Norwell, MA,
1996).
'lq 2 -- [12] J. Dattorro, "The Implementation of Recursive
- - - N (38) Digital Filters for High-Fidelity Audio," J. Audio Eng.
M
Soc., vol. 36, pp. 851-878 (1988 Nov.); Comments,
which is independent of the sample rate, as expected. ibid. (Letters to the Editor), vol. 37, p. 486 (1989 June);
This result [Eq. (38)] indicates that the noise power Comments, ibid. (Letters to the Editor), vol. 38, pp.
spectral level is proportional to the noise power. This 149-151 (1990 Mar.).
means that if the noise power drops by 6 dB, then so [13] P. A. Regalia and S. K. Mitra, "Tunable Digital
will its average spectral level (assuming M fixed). We Frequency Response Equalization Filters," IEEE Trans.
needed to know this to justify the claim (supporting Acoust., Speech, Signal Process., vol. ASSP-35, pp.
fidelity criterion 1 in Section 3.3.2) that the average 118-120 (1987 Jan.).
difference in truncation noise spectral level between a [14] A. V. Oppenheim and R. W. Schafer, Discrete-
16-bit and a 20-bit quantized signal is 24 dB [ = (20 - Time Signal Processing (Prentice-Hall, Englewood
16) x 61. Cliffs, NJ, 1989).
[ 15] J. Strawn, Ed., Digital Audio Signal Processing
4 REFERENCES (A-R Editions, Madison, WI, 1985).
[16] J. A. Moorer, "The Manifold Joys of Conformal
[1] D. C. Andreas, J. Dattorro, and J. W. Mauchly, Mapping: Applications to Digital Filtering in the Stu-
"Digital Signal Processor for Audio Applications," U.S. dio," J. Audio Eng. Soc., vol. 31, pp. 826-841 (1983
patent 5,517,436 (1996 May 14). Nov.).
[2] W. G. Gardner, "Reverberation Algorithms," in [17] D. P. Rossum, "Dynamic Digital IIR Audio Fil-
Applications of Signal Processing to Audio and Acous- ter and Method which Provides Dynamic Digital Filter-
tics, M. Kahrs and K. Brandenburg, Eds. (Kluwer Aca- ing for Audio Signals," U.S. patent 5,170,369 (1992
demic, Norwell, MA, 1997). Dec.).
[3] J. A. Moorer, "About this Reverberation Busi- [18] T. I. Laakso, "Error Feedback for Reduction of
ness," Computer Music J., vol. 3, pp. 13-28 (1979 Quantization Errors due to Arithmetic Operations in Re-
June), also in Foundations of Computer Music, C. Roads cursive Digital Filters," D. Tech. thesis, Rep. 9, Labora-
and J. Strawn, Eds. (MIT Press, Cambridge, MA, tory of Signal Processing and Computer Technology,
1988). Helsinki University of Technology, Espoo, Finland
[4] D. Griesinger, "Practical Processors and Pro- (1991).
grams for Digital Reverberation," in Audio in Digital [19] K. Steiglitz, A Digital Signal Processing Primer
Times, Proc. Audio Eng. Soc. 7th Int. Conf. (Toronto, (Addison-Wesley, Menlo Park, CA, 1996).
Ont., Canada, 1989 May 14-17), pp. 187-195. [20] Moog Voltage-Controlled Ladder Filter, In-
[5] B. A. Blesser and K. O. Bader, "Electric Reverber- ternet: http://www.sara.nl/Rick.Jansen/Emusic/Moog/
ation Apparatus," U.S. patent 4,181,820 (1980 Jan. 1). moogvcf_schematic.gif and Internet: http://www-
[6] D. Griesinger, "Room Impression, Reverberance, ccrma, stanford, edu/-- stilti/papers/moogvcf, pdf.
and Warmth in Rooms and Halls," presented at the 93rd [21] D. Rossum, "Making Digital Filters Sound 'An-
Convention of the Audio Engineering Society, J. Audio alog,'" in Proc. Int. Computer Music Conf. (San Jose,
Eng. Soc. (Abstracts), vol. 40, p. 1064 (1992 Dec.), CA, 1992), pp. 30-33.
preprint 3383. [22] "CEM3328 Four Pole Low Pass VCF," Curtis
[7] M. Schroeder, "Natural Sounding Artificial Re- Electromusic Specialties, Los Gatos, CA (1983).
verberation," J. Audio Eng. Soc., vol. 10, p. 219 [23] H. Chamberlin, Musical Applications of Micro-
(1962 July). processors (Hayden, Indianapolis, IN, 1980).
[8] J. M. Jot and A. Chaigne, "Digital Delay Net- [24] Mathematica, version 2, Wolfram Research,
works for Designing Artificial Reverberators," presented Champaign, IL (1994).
at the 90th Convention of the Audio Engineering Soci- [25] B. L. Evans, L. J. Karam, K. A. West, and
ety, J. Audio Eng. Soc. (Abstracts), vol. 39, p. 383 J. H. McClellan, "Learning Signals and Systems with
(1991 May), preprint 3030. Mathematica," IEEE Trans. Education, vol. 36, pp.
[9] J. O. Smith, "Elimination of Limit Cycles and 72-78 (1993 Feb.).
Overflow Oscillations in Time-Varying Lattice and Lad- [26] K. W. Martin and M. T. Sun, "Adaptive Filters
der Digital Filters," in Music Applications of Digital Suitable for Real-Time Spectral Analysis," IEEE Trans.
Waveguides, Rep. STAN-M-39, Center for Computer Circuits Sys., vol. CAS-33, pp. 218-229 (1986 Feb.);
Research in Music and Acoustics (CCRMA), Dept. of also IEEE J. Solid-State Circuits, vol. SC-21, no. 1
Music, Stanford University, Stanford, CA (1987 (1986 Feb.), pp. 108-119.
May). [27] M. R. Petraglia, S. K. Mitra, and J. Szczupak,
[10] A. V. Oppenheim, Ed., Applications of Digital "Adaptive Sinusoid Detection Using IIR Notch Filters

J. Audio Eng. Soc, Vol. 45, No 9, 1997 September 683

DATTORRO PAPERS

and Multirate Techniques," IEEE Trans. Circuits Sys. [30] T. I. Laakso, P. S. R. Diniz, I. Hartimo, and
H, vol. 41, pp. 7 0 9 - 7 1 7 (1994 Nov.). T. C. Macedo, Jr., "Elimination of Zero-Input and
[28] T. Kwan and K. Martin, "Adaptive Detection Constant-Input Limit Cycles in Single-Quantizer Re-
and Enhancement of Multiple Sinusoids Using a Cascade cursive Filter Structures," IEEE Trans. Circuits Sys. H,
IIR Filter," IEEE Trans. Circuits Sys., vol. 36, pp. vol. 39, pp. 6 3 8 - 6 4 6 (1992 Sept.).
9 3 7 - 9 4 7 (1989 July). [31] A. Luthra, "Extension of Parseval's Relation to
[29] P. P. Vaidyanathan, Multirate Systems and Fil- Nonuniform Sampling," IEEE Trans. Acoust., Speech,
ter Banks (Prentice-Hall, Englewood Cliffs, NJ, 1993). Signal Process., vol. 36 (1988 Dec.), pp. 1909-1911.

THE AUTHOR

Jon Dattorro is from Providence, RI. He trained as a tal signal processing under S. C. Bass. He is currently
classical pianist, attended the New England Conservatory working towards a Ph.D. in electrical engineering at Stan-
of Music where he studied composition and electronic ford University.
music, and performed as soloist with Myron Romanul and He designed the Lexicon Inc. model 2400 Time Com-
the Boston Symphony Orchestra for Children's Concerts pressor with Charles Bagnaschi and Francis F. Lee in
at Symphony Hall. His scores include a ballet and a pi- 1986, and he designed most of the audio effects from
ano concerto. Ensoniq Corp. between 1987 and 1995. He shares two
Mr. Dattorro received a B.S.E.E. with highest distinc- patents in digital signal processing chip design with David
tion from the University of Rhode Island in 1981, where C. Andreas, J. William Mauchly, and Albert J. Charpen-
he was a student of Leland B. Jackson. In 1984 he received tier. Personal mentors are Salvatore J. Fransosi, Pozzi
an M.S.E.E. from Purdue University, specializing in digi- Escot, and Chae T. Goh.

684 J Audio Eng Soc, Vol. 45, No. 9, 1997 September