ArticlePDF Available

Abstract and Figures

An efficient approach to the generation of classical synthesizer waveforms with reduced aliasing is proposed. This paper introduces two new classes of polynomial waveforms that can be differentiated one or more times to obtain an improved version of the sampled sawtooth and triangular signals. The differentiated polynomial waveforms (DPW) extend the previous differentiated parabolic wave method to higher polynomial orders, providing improved alias-suppression. Suitable polynomials of order higher than two can be derived either by analytically integrating a previous lower order polynomial or by solving the polynomial coefficients directly from a set of equations based on constraints. We also show how rectangular waveforms can be easily produced by differentiating a triangular signal. Bandlimited impulse trains can be obtained by differentiating the sawtooth or the rectangular signal. An objective evaluation using masking and hearing threshold models shows that a fourth-order DPW method is perceptually alias-free over the whole register of the grand piano. The proposed methods are applicable in digital implementations of subtractive sound synthesis.
Content may be subject to copyright.
Alias-Suppressed Oscillators Based on Differentiated
Polynomial Waveforms
Vesa Välimäki, Senior Member, IEEE, Juhan Nam, Julius O. Smith, and Jonathan S. Abel
Abstract—An efficient approach to the generation of classical
synthesizer waveforms with reduced aliasing is proposed. This
paper introduces two new classes of polynomial waveforms that
can be differentiated one or more times to obtain an improved
version of the sampled sawtooth and triangular signals. The
differentiated polynomial waveforms (DPW) extend the previous
differentiated parabolic wave method to higher polynomial or-
ders, providing improved alias-suppression. Suitable polynomials
of order higher than two can be derived either by analytically
integrating a previous lower order polynomial or by solving the
polynomial coefficients directly from a set of equations based
on constraints. We also show how rectangular waveforms can be
easily produced by differentiating a triangular signal. Bandlimited
impulse trains can be obtained by differentiating the sawtooth
or the rectangular signal. An objective evaluation using masking
and hearing threshold models shows that a fourth-order DPW
method is perceptually alias-free over the whole register of the
grand piano. The proposed methods are applicable in digital
implementations of subtractive sound synthesis.
Index Terms—Acoustic signal processing, antialiasing, audio os-
cillators, music, signal synthesis.
UBTRACTIVE synthesis is a traditional sound generation
principle used in music synthesizers. Recently, it has be-
come interesting because it is used in digital modeling of analog
sound synthesis. The basic idea in subtractive synthesis is first
to generate a signal with a rich spectral content, and then to filter
that signal with a time-varying resonant filter. The input signal
is usually a periodic classical waveform, such as a sawtooth,
a rectangular, or a triangular wave, or a sum of two or more
such waveforms with different fundamental frequencies. In de-
veloping a digital implementation of subtractive synthesis, the
main challenge is to find computationally efficient algorithms
for generating periodic waveforms that do not suffer from exces-
sive aliasing distortion. This paper discusses a new method for
the generation of the classical waveforms with reduced aliasing.
Manuscript received March 31, 2009; revised June 03, 2009. Current ver-
sion published April 14, 2010. The work of V. Välimäki was supported by the
Academy of Finland under Project 126310. The associate editor coordinating
the review of this manuscript and approving it for publication was Dr. Dan Ellis.
V. Välimäki is with the Department of Signal Processing and Acoustics,
TKK—Helsinki University of Technology, FI-02015 TKK, Espoo, Finland, and
also with the Center for Computer Research in Music and Acoustics (CCRMA),
Stanford University, Stanford, CA 94305 USA (e-mail:
J. Nam, J. O. Smith, and J. S. Abel are with the Center for Computer
Research in Music and Acoustics (CCRMA), Stanford University, Stanford,
CA 94305 USA (e-mail:;;
Digital Object Identifier 10.1109/TASL.2009.2026507
Fig. 1. Discrete-time signals generated by sampling piecewise polynomial pe-
riodic functions. (a) The linear ramp
, or the trivial sawtooth wave-
form, (b) the parabolic
, (c) the cubic , (d) the fourth-order
, (e) the fifth-order , and (f) the sixth-order poly-
nomial function. The fundamental frequency of the signals is 2960 Hz (MIDI
note #102) and the sampling rate is 44.1 kHz.
It is based on differencing a digital signal that is obtained by
sampling a piecewise polynomial waveform.
A straightforward algorithm to generate a discrete-time saw-
tooth wave with signal values between
1 and 1 samples the
continuous-time rising ramp function and shifts it to remove the
dc offset
is the sample index, is the funda-
mental frequency of the tone, and
is the sampling interval.
An example of the signal produced using this method is given
in Fig. 1(a). This bipolar modulo counter can be realized easily
with a computer or a signal processor. Unfortunately, the dis-
crete-time audio signal obtained using (1) is notorious for its
distorted sound quality [1], [2]. The problem is that the contin-
uous-time sawtooth wave is not bandlimited, but it has an infi-
nite number of spectral components falling off at approximately
6 dB per octave. The discretization of the waveform leads to
aliasing where all spectral components above the Nyquist limit
are reflected down to the audible frequency range, when
is the sampling rate, see Fig. 2(a). This is heard as disturbing
An easy way to improve the quality of the trivial sawtooth
wave is to oversample the signal (see [1, pp. 116–117]). The
aliasing resulting from using (1) is less severe when the funda-
mental frequency is small with respect to the sampling rate. In
1558-7916/$26.00 © 2010 IEEE
Fig. 2. Spectra of signals shown in Fig. 1. (a) , (b) , (c)
, (d) , (e) , and (f) . The spectra were computed
from 1-s-long signals using a 262 144-point fast Fourier transform (FFT) with
a Chebyshev window that has its sidelobes at the level of
120 dB. The circles
indicate the desired harmonic components. The dashed line shows in each case
the spectral slope, which is approximately
practice, this approach leads to a highly oversampled and there-
fore inefficient system, because the spectrum of the sampled
sawtooth wave decays slowly [3].
An ideal bandlimited sawtooth wave can be computed using
additive synthesis, in which each harmonic component is gen-
erated separately using a sinusoidal oscillator [4], [5]. The ideal
sawtooth waveform can be expressed as
is the number of harmonics below
the Nyquist limit that will be computed. The drawback of this
approach is that the number of sinusoidal oscillators becomes
large when the fundamental frequency is low. Computational
savings can be obtained by omitting the highest frequencies or
by storing sets of sinusoidal components in several wavetables,
as in group additive synthesis [6]. Nevertheless, this approach
is considerably more costly than (1).
Generation of alias-free waveforms for subtractive synthesis
has received consideration over the years. Winham and Stei-
glitz [7] and Moorer [8] have proposed the use of a closed-form
summation formula to generate a waveform containing a speci-
fied number of sinusoidal components. A low-order digital filter
can shape the roll-off rate of the spectrum so that, for example,
the sawtooth waveform can be closely approximated. While this
method does not require as many operations per sample as ad-
ditive synthesis at low frequencies, it has its drawbacks: it re-
quires a division per sample, and this leads to numerical errors
that must be controlled. One solution is to compute many ver-
sions of the bandlimited sawtooth waveform using (2) with a
different number of harmonic components, and then use them
in wavetable synthesis [4]. This is efficient to compute and free
of numerical problems, but it consumes much memory. A vari-
ation that saves memory consists of a small set wavetables, such
as one per octave, which are played at the various speeds using
sample rate conversion techniques [9].
Stilson and Smith [10], [11] have introduced the idea of
generating alias-free waveforms using a bandlimited impulse
train (BLIT) and filters. This method uses an oversampled
windowed sinc function that is stored in memory. Recently, the
modified frequency modulation technique has been proposed
for synthesizing an approximately bandlimited pulse train [12].
Brandt has proposed to linearly combine a trivial waveform
with a bandlimited pulse to reduce aliasing [13]. The pulse used
is a minimum-phase bandlimited step function (minBLEP)
obtained by integrating a minimum-phase windowed sampled
sinc function [13]. The computational load of the BLEP-based
algorithm depends on the fundamental frequency. The algo-
rithm consumes some memory, because an oversampled BLEP
residual sequence must be designed and stored in advance for
real-time implementation. The BLEP residual is the difference
of an ideal step function and the BLEP approximation [2].
Recently, it has been shown that the BLEP residual sequence
can be obtained by sampling a low-order polynomial curve
[2]. The table lookup is not needed then and the oversampling
factor is replaced with a fractional delay variable, which can
vary continuously.
et al. introduced a different approach to derive an-
tialiasing oscillator algorithms. They full-wave rectify a sine
wave and modify the resulting signal with digital filters [14].
Classical waveform approximations computed with Lane’s
algorithm contain some aliased components, but they are
suppressed in comparison to the trivial waveforms. Pekonen
and Välimäki used recently highpass and comb filtering as a
postprocessing method to suppress aliasing in waveforms [15].
It has been shown that a signal that closely approximates the
sawtooth wave but having less aliasing can be produced by dif-
ferentiating a piecewise parabolic waveform [16]. The simplest
version of this algorithm generates the bandlimited sawtooth
waveform approximation in four stages: First a trivial sawtooth
waveform is generated using the bipolar modulo counter (1),
then the signal is raised to the second power, it is differenti-
ated with a first difference filter with transfer function
and, finally, the obtained waveform is scaled. A two-times over-
sampled version of the differentiated parabolic wave algorithm
yields improved alias suppression and leaves room for optimiza-
tion by using various choices of decimation filters [16], [2].
In this paper, we propose an extension to the differentiated
parabolic wave method mentioned above. Our ultimate design
goal is not to suppress the aliasing completely, but to reduce
it sufficiently at selected frequency areas so that it will be in-
audible. We take advantage of the frequency-dependent sensi-
tivity and the frequency masking phenomenon of the human au-
ditory system [17]. Other goals are computational efficiency and
small memory space requirement. In this paper we introduce a
new class of polynomial waveforms that must be differentiated
more than once to obtain an improved version of the sampled
sawtooth signal. Suitable polynomials of order higher than two
can be derived either by analytically integrating a previous lower
order polynomial or by solving the polynomial coefficients di-
rectly from a set of equations based on constraints. A multirate
version of the algorithm oversamples and decimates the polyno-
mial signal with a simple filter prior to differentiation. Another
new class of polynomial signals can be differentiated to produce
alias-suppressed triangular signals.
The basics of the new approach are introduced and a multirate
method is considered in Section II. Section III suggests algo-
rithms to synthesize triangular, rectangular, and impulse train
signals. The audio quality of the produced waveforms is ana-
lyzed and is compared against previous methods in Section IV.
Section V concludes this paper.
In this section, we show that it is possible to extend the
principle used in the differentiated parabolic wave method
for the generation of sawtooth waveforms to polynomials of
order higher than two. The motivation to use a high-order
piecewise polynomial signal with polynomial order
is that
its spectrum decays about
dB per octave. For comparison,
the spectrum of the sawtooth wave falls off only about
per octave. The faster roll-off rate leads to suppressed aliasing
[16], [18], [2]. Multiple differentiations can restore the desired
spectral tilt for the harmonic components. In the following, it
is shown how the desired polynomials can be derived and how
the resulting signal can be scaled. A multirate implementation
is also discussed.
A. Derivation of Polynomials Using Analytical Integration
It was shown in a previous work [16] that a digital signal that
closely resembles the sawtooth wave, but with a smaller aliasing
error, can be generated by differentiating a sampled, piecewise
parabolic waveform. The piecewise parabolic waveform can be
obtained by integrating one period of the sawtooth waveform,
which is a linear function within the period, that is
Scaling can be dealt with separately, so we may use the poly-
for synthesis. Fig. 1(b) shows the sampled parabolic
waveform, and Fig. 2(b) shows its spectrum, which decays about
12 dB per octave.
In principle, differentiating (3) restores the piecewise linear
function. However, when a discrete-time signal consisting of
samples taken from the second-order polynomial curve is differ-
entiated with a digital filter, the result is different from a trivial
sawtooth waveform: one or more samples near the joint of two
contiguous periods will be different. This is seen in Fig. 3(b),
which shows the resulting discrete-time signal when a first-order
has been applied. A single modified sample
forms a smooth transition from one period to the next one. This
time-domain property of the differentiated polynomial wave-
forms (DPW) waveform reduces aliasing. Fig. 4(b) shows the
spectrum of the second-order DPW sawtooth waveform, which
Fig. 3. Alias-reduced sawtooth waveforms generated by differencing polyno-
mial periodic functions
times. (a) The trivial sawtooth waveform, (b) the
parabolic, (c) the cubic, (d) the fourth-order, (e) the fifth-order, and (f) the sixth-
order DPW signals.
Fig. 4. Spectra of the sawtooth waveforms shown in Fig. 3. (a) The trivial saw-
tooth waveform, (b) the parabolic, (c) the cubic, (d) the fourth-order, (e) the
fifth-order, and (f) the sixth-order polynomial function. The parameters used in
the spectral analysis are the same as in Fig. 2. The circles indicate the desired
harmonic components. The dashed line shows the ideal spectral slope of the
sawtooth signal, which is approximately
6 dB/octave.
has generally weaker aliased components than the spectrum of
the trivial sawtooth waveform, cf. Fig. 4(a).
We can continue integrating (3), but an arbitrary constant of
must be assumed:
The constant
can be determined by the reasonable restriction
that function
must not contain a discontinuity when re-
peating, that is,
. We then obtain ,
and thus . Again, we may skip the scaling
here and use
instead. Fig. 1(c) plots the sampled
version of this polynomial signal.
Fig. 2(c) shows that the spectrum of the cubic waveform has
about 6 dB per octave steeper slope than the parabolic wave-
form. This leads to further reduction of aliasing. Filtering twice
with a first-order difference yields a sawtooth-like signal, which
has a transition region of two samples between each period, see
Fig. 3(c). The corresponding spectrum is shown in Fig. 4(c). We
also see that the level of harmonics at high frequencies above
10 kHz is several dB too low. This is caused by the poor ap-
proximation of the differentiation by a first-order finite differ-
ence. We have found that it is possible to compensate for this
approximation error using a second-order IIR equalizer, if de-
sired. Then the level of harmonics will not deviate more than a
fraction of a decibel in the audio range.
By integrating the cubic polynomial, we obtain the fourth-
order polynomial, and so on. However, we will next show how
to derive the polynomial functions directly.
B. Derivation of Polynomials by Solving Linear Equations
In this section, we show that it is possible to directly solve
for the coefficients of an
th-order polynomial from constraints
that describe certain desired characteristics of the polynomials.
denote the
th-order polynomial. The coefficients are uncon-
strained at this point. Without loss of generality, we can let the
polynomial be monic (i.e.,
), because a scaling factor
can separately deal with the overall level. The key idea in the
class of techniques that we propose is to impose a maximally
flat wrap-around at the switch-back point in the sawtooth signal.
ranges from 1 to 1, and we do not want a constant
term for this first-order polynomial, so we can set
There remain
coefficients that we
may choose so as to maximize spectral roll-off rate in the signal
(cf. (1)
As is well known in the context of windows used for Fourier
analysis, spectral roll-off rate is maximized by means of “max-
imally flat” tapering at the discontinuity points (window end-
points). Specifically, a maximally flat “splice” of the function
is obtained at by matching the maximum number of
leading terms in the Taylor expansion of
about to zero. In
the present context, we wish to equate as many leading deriva-
tives of
as possible at the “wrap-around points” .
That is, we desire
and so on, for as many derivatives as possible. (Note that cubic
splines, being third-order polynomials, match both function
value and slope at each spline junction, or “knot.”)
Even and odd polynomials are of interest in this work, be-
cause the derivative of an even-order polynomial is odd, and
vice versa. Furthermore, every polynomial can be written as the
sum of an even and odd polynomial. Every even polynomial sat-
isfies a desired property that
, and the same is
also true for every even-order derivative of
. Since every other
coefficient of an even polynomial is zero, the number of odd
derivatives is comparable to the number of coefficients. Thus,
substantially all degrees of freedom in an even polynomial are
devoted to satisfying the maximally flat constraints we seek.
For odd polynomials, the only solution for
is the zero solution, so they must have the property that their
coefficients sum to zero, which leads to
. Since is
odd, we obtain
automatically. Any even number
of derivatives of an odd polynomial is odd, so similar remarks
apply. An odd number of derivatives gives an even polynomial,
which is spontaneously maximally flat at the wrapping points,
as discussed in the previous paragraph.
Choosing an even or odd starting polynomial—depending on
is even or odd—constrains every second coefficient
to zero. Additionally, we can argue that
can be zero: in the
even case, it is arbitrary, and in the odd case, it must be zero, so
it may as well be zero in every case.
For even
, we have a polynomial of the form
already no matter what coefficient values
are chosen. For odd
, we have a polynomial of the form
and we must have
Going back to the even case, to obtain our first constraint, we
must differentiate and equate the coefficient sum to zero. Then
we differentiate twice more, equate the sum of coefficients to
zero, and so on, until we get down to order 3 which determines
(because ). Finally, we back-solve the previous
equations in reverse order to obtain all coefficients. In matrix
form, it is a simple upper triangular system with integer en-
tries made from every other derivative. Matrix inversion is not
needed, as the equations can be solved starting from the bottom.
The odd case goes the same way, except that we start with
instead of its derivative, i.e., the coefficients of must
sum to zero, as well as those for
, and so on down
to order 2, where
is again determined and the rest can be
found by back-solving.
With the above constraints, it is possible to derive a polyno-
mial function for any
. A few example cases are shown next.
1) Examples: The second-order case is of the form
, and
there are no other terms or coefficients, because we must have
(to keep it even) and we can choose (no effect),
that is
This case has been known for several years from an earlier work
[16], where it was derived by integrating a single period of a
linear ramp function, as in (3).
1 TO
For the third-order case , the form is .
is odd (making the polynomial odd), these coefficients
must sum to zero. This yields
, and the cubic polynomial
This is equivalent to the third-order function that we obtained
previously with the integration method in Section II-A
, the form is . Differentiating
with respect to
yields . Setting
yields . The fourth-order polynomial is thus
, the form is yielding
for our first constraint. Next, we need the first and second
and .
yields , and back-solving gives
. The fifth-order polynomial is then
, we start with and go
, giving ,or
for (1). Next, and
, which gives ,
and back-solving gives
. The sixth-order
polynomial is therefore
Similarly, it is possible to derive a polynomial of any order
for this purpose. We will use the above polynomials of orders
1 to 6 in the examples of this paper. They are also included in
Table I. For practical implementations, we suggest to use the
first-order difference
as an approximate for derivative,
as was done previously in the second-order case [16].
C. Scaling Factors
When the linear ramp function is raised to a high power
and differenced several times, the signal level becomes small.
For this reason, signal scaling is a crucial part of the synthesis
method. We derive next the scaling factors needed for different
orders of polynomials.
1) Continuous Polynomial Scaling: When an
monic polynomial, such as (5) with
, is differentiated
times, it becomes a first-order polynomial of the form
, where denotes the factorial of , and is zero
by construction in our case (since we employ only even or odd
polynomials). Therefore, the scale factor for ideal differentia-
tors is
2) Discrete Waveform-Preserving Scaling: When the ideal
differentiator is replaced by a first-order finite difference, we
have instead
, where denotes
the first-order difference operator, and
is defined in (6).
If there is no modulo operation between time-steps
and ,
then the highest-order term in
is found to be .
Thus, the
-order term is canceled and the polynomial order
is reduced by one. After
applications of the difference
operator, we obtain
It can be shown that
corresponds to a half-sample delay
per application of
, which we will ignore (since it is easily
compensated by a time shift). Thus, our waveform-preserving
scaling factor for the finite-difference case becomes
denotes the waveform period in samples. The
scaling factors of this form are presented in Table II, second
column, for
to . Notice that the scaling factors are
proportional to
and inversely proportional to the funda-
mental frequency
raised to the power . The scaling
factors also contain a large integer in the denominator. To avoid
the division, when the fundamental frequency is changed, the
scaling factors can be precomputed and stored in a table.
3) Fundamental-Matching Scaling: The previously derived
scale factor for differenced polynomials exactly preserves the
sawtooth waveform at the sample points over regions not in-
cluding the modulo discontinuity. While some applications may
prefer this type of scaling, it is more typical in audio applica-
tions to optimize the scaling according to psychoacoustic cri-
teria. Since a scale factor offers only one degree of freedom, it
can be argued that the best use for it is to maintain correct ampli-
tude at the fundamental frequency, especially as that frequency
is changed dynamically. Thus, we derive a scale factor that can-
cels the frequency-dependent difference between the ideal dif-
ferentiator and the first-difference filter at the fundamental.
The frequency response of an ideal differentiator is
that of a first-order difference is
, where
is radian frequency, is the imaginary unit, and denotes
the sampling interval. Neglecting the linear phase term corre-
sponding to a half-sample delay, correcting the amplitude at
is accomplished by the scale factor
This scaling factor is included in the third column of Table II for
each polynomial order
The scaling factors according to (17) are displayed in
Fig. 5(left) for polynomial orders
to . It is seen that
the gain is very large, when the fundamental frequency is low.
For example, for
, the required scaling factor for the
lowest note on the piano (
Hz) is , or about
233 dB. Implementing a multiplication with a large number like
this requires special care in finite-accuracy arithmetic systems,
such as in 16-bit fixed-point signal processors. A solution is to
implement the scaling in different stages: The first part of the
scaling factor can be used after evaluating the polynomial and
the rest between the first-difference filters, which reduce the
amplitude of low-frequency signals. In practice, the multistage
scaling means that more multiplications than just one must be
used, though all but one of them can be approximated by the
nearest power of 2.
Using the scaling factors presented in Table II, the amplitude
of the sawtooth signals generated using the DPW oscillator can
be limited not to exceed 1.0. This can be seen in Fig. 5(right),
where the amplitude of sawtooth wave signals with 88 funda-
mental frequencies ranging from 27.5 Hz to 4186 Hz is shown.
The amplitude has been estimated as the absolute maximum of
sample values from a 1-s-long signal. For high fundamental fre-
quencies, the amplitude is less than one, or 0 dB, but no more
than about 2.5 dB less.
4) Bounded Scaling: We may alternatively solve for the
scaling factors that are needed to maintain a maximum value
of one. We again assume that the first difference operation is
used to approximate differentiation. In the case of the parabolic
Fig. 5. (left) Gain factors of the differentiated polynomial waveforms of orders
, see Table II, third column, and (right) amplitude of the sawtooth
waveforms using these scaling factors. For
, the first octave is opted
out in the right figure, since the scaling factor is very large. Fundamental fre-
quencies between 27.5 Hz (MIDI note #21) and 4186 Hz (MIDI note #108) are
polynomial of (10), this can be done by requiring that the scaled
difference of the largest and second largest sample values is
1.0, that is
is the scaling factor. We can solve as
As mentioned in [18], this can well be replaced with a sim-
plified version
, which is also the factor that we ob-
tained previously using another method, as seen in Table II. This
approximation deviates from (19) by less than 1 dB, when the
fundamental frequency is below 4.0 kHz and the sampling fre-
quency is 44.1 kHz.
Instead of deriving exact polynomial formulas for the scaling
factors of high-order methods in this way, we propose to use
simple approximations shown in Table II.
D. Computational Complexity
Next we discuss the computational load of the proposed
methods. Table I presents the number of operations needed for
different parts of the algorithms, such as evaluating the polyno-
mials and scaling. It is seen that the number of multiplications,
including scaling, is about
. The total number of operations
turns out to be about
, where is the polynomial order. It
is assumed that the first-difference filter is used to implement
all derivatives.
Fig. 6. Spectra of the sawtooth waveforms generated using the two times over-
sampled method, where a two-point averager is used as the decimation filter. (a)
The trivial sawtooth waveform, (b) the parabolic, (c) the cubic, (d) the fourth-
order, (e) the fifth-order, and (f) the sixth-order polynomial function. The circles
indicate the desired harmonic components, and the dashed line is the spectral
envelope of the ideal sawtooth waveform.
E. Multirate Method
It has been shown previously that an oversampled version
of the second-order DPW algorithm can improve the alias-re-
duction at low frequencies [16], [2]. The multirate version of
the algorithm oversamples and decimates the polynomial signal
with a first-order filter prior to differentiation. It is expected that
this idea should perform very well also when third or higher
order polynomials are used. The steepened spectral roll-off rate,
which is about
18 dB per octave for and 6 dB more
for each higher order, helps to further reduce aliasing.
Fig. 6 shows the spectra of sawtooth signals produced using
polynomials of order
to with two times oversampling.
A two-point averaging filter is used for decimating the signal,
as proposed in [16]. Fig. 6 should be compared against Fig. 4
to see whether the improvements are useful. It can be noticed
that the aliasing is reduced in all cases at low frequencies, par-
ticularly below the first harmonic, which is at 2960 Hz in this
case. The improvement gets better as
increases. In fact, the
oversampled method of order
[see Fig. 6(d)] is practi-
cally as good as the nonoversampled method of order
[see Fig. 4(e)].
However, since the computational complexity is also dou-
bled, when two samples must be generated in each sampling in-
terval, it is impractical to use the oversampling method. Notice
that according to Table I, increasing the polynomial order gener-
ally increases the computational burden less than oversampling.
An exception is when
is increased from 2 to 3, since then the
number of operations is exactly two times larger. However, the
regular method with
is still better than the oversam-
pled method with
, so oversampling does not give any
advantage even in this case. The DPW method is itself a very
efficient signal processing method, which cannot be improved
with a simple trick.
Fig. 7. Algorithms to generate (a) the trivial triangular, (b) the differentiated
bipolar parabolic
, and (c) the differentiated cubic triangular
waveforms. The modulo counter produces a trivial sawtooth signal with values
and 1. The block containing full-wave rectifies its input signal.
In the section, we will show how to realize various classical
waveforms using higher-order DPW methods. These include the
triangular and square waveforms and the bandlimited impulse
A. Triangular Wave
The trivial triangular waveform can be obtained by sampling
a function that goes linearly from
to and linearly back
in each period. This waveform can be generated from the
bipolar module counter (1) in the following way:
The block diagram of this algorithm is shown in Fig. 7(a).
It has been shown in [18] that differentiating a bipolar par-
abolic waveform leads to an alias-suppressed triangular signal.
The bipolar parabolic waveform is obtained as
Here the term
corresponds to the fullwave recti-
fied sawtooth signal, which is a triangular signal with a dc offset.
Note that it deviates from (20) because of its dc level. The mul-
tiplication of the bipolar modulo counter signal and its shifted,
full-wave-rectified version leads to a piecewise parabolic wave-
form that has its every second cycle inverted (nonpositive), see
Fig. 8(b). Differentiation and scaling produces the alias-sup-
pressed triangular signal:
Depicted in Fig. 7(b), this new algorithm is simpler than the
one presented earlier by Välimäki and Huovilainen [18], but
both algorithms produce the same output signal.
Alias-suppressed triangular signals with differentiated high-
order polynomial signals can be derived as follows. The poly-
nomial signals can be obtained by applying a polynomial func-
to the trivial sawtooth wave (1). Each polynomial func-
of order must be an odd function between points
Fig. 8. Signals generated by sampling periodic polynomial functions used for
producing triangular signals. (a) The bipolar module counter (trivial sawtooth
waveform), (b) the bipolar parabolic
, (c) the cubic , (d) the
, (e) the fifth-order , and (f) the sixth-order
polynomial signals.
Fig. 9. (a) Trivial triangular and alias-reduced triangular waveforms of orders
, (c)
, (d)
, (e)
, and (f)
generated by
differencing polynomial functions of Fig. 8
and . Additionally, the polynomial must have
a special property that its two negative and positive quadrants
are symmetric with each other with respect to the points
and , that is, ,
because the waveform must contain only odd harmonics of the
fundamental. The continuous-time triangular waveform itself
fulfills these requirements.
We use monic polynomials and skip the constant terms, as
was previously done in Section III. Interestingly, we must now
include fullwave rectified terms
, like in (20) and (21). Note
that even powers of
, such as and , are even functions
and , and thus we must replace them with
and , respectively. We may again investigate even and odd
polynomial functions separately, as in Section III. For even
we use a polynomial function of the following general form:
For odd
, we use the following form:
Values of coefficients
can be derived for any by ap-
plying continuity conditions for the function itself and its deriva-
tives. We match the derivatives from each side at
, that is
and so on, where
means that zero is approached from the
positive side. In addition, for odd-order polynomials we match
the derivatives at
using the following conditions:
and so on. Since the polynomial functions are symmetric about
the point
, the above conditions are always true for even-
order derivatives. Therefore, for odd-order derivatives only the
following conditions are necessary:
and so on.
Based on the matching of derivatives at each side of zero, the
terms of the form
, when is even, must have a weight
zero. The reason is that the sign of the derivative is different on
each size, and therefore the only way to match the term is to set
their weighting coefficient to zero.
We can verify that (19) uses a second-order polynomial func-
of the form that satisfies
the derivative matching at
. The derivative is .
Evaluating it at
gives . Setting yields
. Thus the second-order polynomial function is
Fig. 8(b) presents the waveform obtained by inserting the
bipolar modulo counter signal into this polynomial function.
Differentiation yields the triangular signal approximation
shown Fig. 9(b). Fig. 7(b) shows the block diagram of the
whole algorithm.
, the form is .We
need to compute the first and second derivatives of the function
on the positive and negative side of zero to see that
must be
1 TO
Fig. 10. Spectra of the triangular waveforms shown in Fig. 9. (a) The trivial
triangular waveform and (b) the second-order, (c) the third-order, (d) the fourth-
order, (e) the fifth-order, and (f) the sixth-order approximations. The parameters
used in the spectral analysis are the same as in Fig. 2. The dashed line shows the
spectral roll-off for the ideal triangular signal, which is about
12 dB/octave.
zero. We then set the derivative to zero at .
This yields
, and the cubic polynomial becomes
This polynomial function also requires a special input signal,
because it only works correctly between
and . There-
fore, the input signal must repeat the same signal values in the
first and the fourth quarter of the period. An appropriate input
signal can be obtained by scaling (18) to have only values be-
and :
The block diagram of the third-order triangular algorithm,
including the input signal generation, is presented in Fig. 7(c).
The cubic waveform and the triangular signal approximation
resulting after differencing twice are shown in Figs. 8(c) and
9(c), respectively.
, the form is .
From its second derivative on both the positive and negative side
of zero, we see that
. We the first derivative
to zero at . This yields . We can then solve
by setting the third derivative to zero at . We get
. The fourth-order polynomial is
The fourth-order waveform and its three times differenced ver-
sion are presented in Figs. 8(d) and 9(d), respectively.
For the case
and for any larger , we can similarly dif-
ferentiate the general polynomial function, differentiate it mul-
tiple times and check the value on each side of points
to solve the coefficients. We skip the lengthy deriva-
tions of cases
and . The fifth- and sixth-order
polynomial waveforms and the four and five times differenced
waveforms are shows in Fig. 9(e), (f), 10(e), and (f), respec-
The sample values of DPW triangular waveforms can be
maintained between
1 and 1 by multiplying the signals with
, that is, just two times larger scaling factors than the DPW
sawtooth signals discussed previously.
We have gathered the polynomial functions for the cases
to in Table III. The third column in Table III shows the input
signal for each case. Notice that when
and ,a
trivial triangular input signal must be used. Comparison against
Table I reveals that the triangular algorithms generally need a
few more operations per sample than the sawtooth algorithms.
However, since the spectral roll-off rate of the triangular wave-
forms is steeper, it is expected that one of the low-order differ-
entiated polynomial algorithms will be sufficient in practice.
Fig. 10 shows the spectra of the triangular waveforms of
Fig. 9. It can be seen that the signals obtained by differencing
high-order polynomial signals suffer from less aliasing than the
trivial triangular signal [Fig. 10(a)]. The quality improves as
the polynomial order is increased. When
or larger, the
aliasing components below the first harmonic are suppressed
more than 100 dB in this case. Notice also that in Fig. 10 the
level of upper harmonics is always below the ideal spectral
envelope that decays
12 db per octave. This is a consequence
of the simple first-order difference used.
B. Rectangular Wave
The rectangular waveform with a 50% duty cycle, i.e., the
square waveform, can be produced easily by differentiating a
triangular signal. For the first-order case, this leads to the algo-
rithm depicted in Fig. 11. The bipolar modulo counter signal is
Fig. 11. First-order rectangular pulse generator based on differentiating the
trivial triangular waveform.
Fig. 12. Alias-reduced square waveforms generated by differencing the trian-
gular signals Fig. 8. (a) The first-order, (b) the second-order, (c) the third-order,
(d) the fourth-order, (e) the fifth-order, and (f) the sixth-order square waveform
full-wave rectified, inverted, shifted, scaled, and finally differ-
enced. An example of the waveform and its spectrum are shown
in Fig. 12(a) and Fig. 13(a), respectively. This waveform has a
transition of one sample at each edge, so this is an improved
version of the trivial square waveform.
Higher-order approximations for the square signal are ob-
tained by differencing the higher-order triangular signals one
more time. Fig. 12 shows the differenced versions of signals in
Fig. 9. Notice how the transition region becomes smoother when
is increased. The corresponding spectra are shown in Fig. 13.
In comparison to Fig. 10, the aliasing has been more suppressed
at low frequencies in every case.
It has been shown in [18] that rectangular waveforms with a
different duty cycle can be produced with the help of sawtooth
signals [19], [18]. One can either subtract the output signals of
two sawtooth oscillators with an appropriate phase shift (e.g.,
half a period for the 50% duty cycle) [18] or filter a sawtooth
signal with a feedforward comb filter [19], which must use an
interpolated delay line. Both methods facilitate also pulse-width
modulation, which is a common technique in subtractive syn-
thesis. When a high-order DPW sawtooth algorithm is used, it
is more cost efficient to produce the rectangular pulses using an
interpolated delay line method rather than running two DPW al-
gorithms in parallel.
C. Bandlimited Impulse Trains
It is known that integrating an impulse train yields a saw-
tooth wave [10], [11]. Therefore, it is obvious that differenti-
ation of the sawtooth waveform yields the impulse train. It is
Fig. 13. Spectra of square wave approximations of Fig. 12. The dashed line
shows the spectral roll-off for the ideal square signal, which is approximately
6 dB/octave.
Fig. 14. (a) Second-order DPW sawtooth and (b) a first-order DPW square
wave. (c) A unipolar and (d) a bipolar bandlimited impulse train obtained by
differencing the signals (a) and (b), respectively.
quite surprising, however, that when a sawtooth waveform ob-
tained with the DPW method is differentiated, the result is an
alias-suppressed pulse train, which can be obtained by inserting
impulses into a B-spline filter. The theory of generating an-
tialiasing impulse trains using interpolation and fractional-delay
filter methods is discussed in a companion paper [20]. Differen-
tiating a DPW square signal yields the bipolar BLIT sequence,
in which there are two pulses in each period and every second
of them is inverted.
Fig. 14 shows two examples, where a second-order DPW
sawtooth [Fig. 14(a)] and the first-order DPW square wave
[Fig. 14(b)] are differentiated with the first-order difference.
Fig. 14(c) is the unipolar BLIT sequence obtained by dif-
ferencing the sawtooth wave. Each pulse consists of two
samples, which are also equivalent to the impulse response
of the linear-interpolation FIR filter, which corresponds to
the first-order B-spline interpolation. Additionally, this BLIT
sequence contains a dc component. Fig. 14(d) shows the differ-
enced square wave, which is a bipolar BLIT signal. Also these
pulses can be obtained using the linear-interpolation filter.
While this observation may not be useful for generating
BLITs, which can be synthesized directly without difference
operations, it shows that the DPW methods and polynomial
interpolation methods are closely related. This has not been
known previously. The aliasing behavior of DPW and BLIT
signals is also easily understood by knowing how they are
related to each other.
In this section, we evaluate the audio quality of signals pro-
duced using the proposed DPW methods. It is of interest to
understand how the quality is improved, when the polynomial
order is increased. We also include in the evaluation the over-
sampled DPW algorithms and compare the results against a pre-
vious algorithm.
Previously, the sound quality of alias-suppressed signals has
been assessed using various signal-to-noise (SNR) like mea-
sures, such as the power ratio of desired and undesired (i.e.,
aliased) components and its F-weighted version [16]. Välimäki
and Huovilainen applied the noise-to-mask ratio (NMR), which
is commonly used for evaluating the quality of perceptual audio
codecs [2]. Timoney
et al. introduced the use of Perceptual Eval-
uation of Audio Quality (PEAQ) for this purpose [12]. None
of these measures is perfectly suited for evaluating the quality
of harmonic tones with tonal disturbance, such as the classical
waveforms with aliasing. We believe that a method that would
account for the masking of the aliasing components by the par-
tials would be the ideal method. This way it would be possible
to decide case by case whether a synthesis method yields a per-
ceptually alias-free output at a given fundamental frequency. We
attempt to develop and use such method in this work.
We compute the FFT spectrum of each tone from a 1-s-long
signal segment using a Hamming window, which facilitates the
scaling of the signal power. We assume that the signals are
played at the 96 dB SPL (sound pressure level). The signals are
scaled so that their power is equivalent to the power of a sine
wave played at 96 dB SPL.
Since the absolute signal level is known after scaling, we can
use a hearing threshold function [21] to find which spectral com-
ponents are inaudible. Additionally, we use the following asym-
metric spreading function model for the masking effect of each
partial on the dB scale [22]
is frequency in Bark units [17], is the level of the
masker (in dB),
is the frequency difference of the masker
and the maskee, and
is a step function that is equal to zero
and otherwise. A masking curve model of (32)
is allocated for each partial of the ideal waveform and is shifted
down by 10 dB from the partials level. We combine the masking
curves of each partial with the hearing threshold function using
the max operation in the frequency domain.
Fig. 15. Spectra (solid line) and masking curves (dashed line) of the DPW saw-
tooth signals of Fig. 3.
Fig. 15 shows examples of the spectra of sawtooth signals
produced using the trivial method and the DPW methods of or-
ders 2 to 6 with their overall masking curves (dashed lines). The
dashed line indicates the overall masking curve for each signal.
It is seen that at this fairly high fundamental frequency, many
aliased components are above the masking curve in Fig. 15(a)
and (b), but in all other cases only five harmonic components
appear above the curve. This implies that aliasing should be au-
dible in the trivial sawtooth [Fig. 15(a)] and in the sawtooth gen-
erated with the DPW with
[Fig. 15(b)]. For higher poly-
nomial orders, the aliasing is completely hidden by the auditory
masking phenomenon or by the disability of the human hearing
to sense very quiet sounds.
Using this approach, we evaluated the sawtooth signals gen-
erated with the DPW methods of orders 1 to 6 for each fun-
damental frequency of the tempered scale (
Hz). We
checked the highest fundamental frequency for which all aliased
components remained below the masking curve. The results are
presented in Table IV, second column. It can be noticed that in-
creasing order
improves the alias-suppression considerably
in every case. We repeated the same for the two times oversam-
pled DPW algorithm (with a two-tap averager as the decimation
filter), see Table IV, third column. The oversampling improves
the performance in most cases, but, as discussed in Section II-E,
it cannot be recommended, since it also doubles the computa-
tional burden.
For comparison, according to our evaluation, a sawtooth
signal produced using the method proposed by Lane et al. [14]
has the highest alias-free frequency of 600.0 Hz—the same
result as with the second-order DPW method. (We used an
optimized version of Lane’s method in which the low-pass filter
is opted out.) However, Lane’s method has been estimated to
have a larger computational cost than the second-order DPW
method [16].
The highest tone of the piano, a C8, with the fundamental
frequency of 4186 Hz, is often considered the highest pitch used
in music. Thus, according to the results of Table I, the fourth-
order DPW method provides a sufficient alias-suppression over
TO 6. T
44.1 kHz
88.2 kHz (O
Fig. 16. Proposed implementation structure in which tones with a low fun-
damental frequency (below
) are produced with a low-order DPW algorithm
) and high fundamental frequencies are produced with a high-order
DPW method (e.g.,
). A unit delay is needed to synchronize the signals.
the normally used range of fundamental frequencies, when the
sample rate is 44.1 kHz. It would be superfluous to use the fifth
or sixth-order DPW methods in that case.
A practical implementation structure for DPW sawtooth syn-
thesis is presented in Fig. 16, where low-frequency tones are
produced by a second-order DPW method and high frequencies
are produced using a fourth-order method. The crossover fre-
may be about 500 Hz. Using a combination of two
DPW algorithms it is possible to avoid a very large scale factor
at low fundamental frequencies but still maintain a good audio
quality at high fundamental frequencies.
V. C
This paper introduced novel classes of polynomial wave-
forms that can be differentiated one or more times to obtain
an alias-suppressed sawtooth or triangular signal. The pro-
posed differentiated polynomial waveform method extends the
previous differentiated parabolic waveform method to higher
orders than two. Square waves can be synthesized by applying
the first-order difference operation to the triangular waveforms.
Additionally, it was observed that the DPW methods are closely
related to polynomial interpolation techniques, because dif-
ferencing a DPW sawtooth or square signal yields a unipolar
or a bipolar pulse train, respectively, consisting of impulse
responses of low-order interpolation FIR filters.
Perceptual evaluation of DPW sawtooth signals shows that
aliasing may be audible in the previously introduced second-
order DPW method at fundamental frequencies above 600 Hz,
when the sampling rate is 44 100 Hz. The fourth-order DPW
method is alias-free up to about 4.6 kHz and thus permits the
synthesis of practically all fundamental frequencies commonly
used in music.
The first author would like to thank A. Huovilainen for
helpful discussions. The second author would like to thank
Prof. M. Bosi for helpful discussions on auditory masking.
[1] H. Chamberlin, Musical Applications of Microprocessors, 2nd ed.
New York: Hayden, 1985.
[2] V. Välimäki and A. Huovilainen, “Antialiasing oscillators in subtrac-
tive synthesis, IEEE Signal Process. Mag., vol. 24, no. 2, pp. 116–125,
Mar. 2007.
[3] M. Puckette, The Theory and Technique of Electronic Music. Singa-
pore: World Scientific, 2007.
[4] F. R. Moore, Elements of Computer Music. Englewood Cliffs, NJ:
Prentice-Hall, 1990, pp. 270–271.
[5] A. Chaudhary, “Bandlimited simulation of analog synthesizer modules
by additive synthesis, in Proc. Audio Eng. Soc. 105th Conv., San Fran-
cisco, CA, 1998, paper no. 4779.
[6] P. Kleczkowski, “Group additive synthesis, Comput. Music J., vol. 13,
no. 1, pp. 12–20, 1989.
[7] G. Winham and K. Steiglitz, “Input generators for digital sound syn-
thesis,J. Acoust. Soc. Amer., vol. 47, pp. 665–666, Feb. 1970, part 2.
[8] J. A. Moorer, “The synthesis of complex audio spectra by means of
discrete summation formulae, J. Audio Eng. Soc., vol. 24, no. 9, pp.
717–727, Nov. 1976.
[9] D. C. Massie, “Wavetable sampling synthesis,” in Applications of Dig-
ital Signal Processing to Audio and Acoustics, M. Kahrs and K. Bran-
denburg, Eds. Norwell, MA: Kluwer, 1998, pp. 311–341.
[10] T. Stilson and J. Smith, “Alias-free digital synthesis of classic analog
waveforms, in Proc. Int. Computer Music Conf., Hong Kong, China,
1996, pp. 332–335.
[11] T. Stilson, “Efficiently-variable non-oversampling algorithms in vir-
tual-analog music synthesis—A root-locus perspective” Ph.D. disser-
tation, Dept. of Elect. Eng., Stanford Univ., Stanford, CA, Jun. 2006
[Online]. Available:
[12] J. Timoney, V. Lazzarini, and T. Lysaght, “A modified FM synthesis
approach to bandlimited signal generation, in Proc. 11th Int. Conf.
Digital Audio Effects, Espoo, Finland, Sep. 2008, pp. 27–33.
[13] E. Brandt, “Hard sync without aliasing, in Proc. Int. Comput. Music
Conf., Havana, Cuba, 2001, pp. 365–368.
[14] J. Lane, D. Hoory, E. Martinez, and P. Wang, “Modeling analog syn-
thesis with DSPs,Comput. Music J., vol. 21, no. 4, pp. 23–41, 1997.
[15] J. Pekonen and V. Välimäki, “Filter-based alias reduction in classical
waveform synthesis,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal
Process., Las Vegas, NV, Mar. 2008, pp. 133–136.
[16] V. Välimäki, “Discrete-time synthesis of the sawtooth waveform
with reduced aliasing, IEEE Signal Process. Lett., vol. 12, no. 3, pp.
214–217, Mar. 2005.
[17] E. Zwicker and H. Fastl, Psychoacoustics: Facts and Models.New
York: Springer-Verlag, 1990.
[18] V. Välimäki and A. Huovilainen, “Oscillator and filter algorithms for
virtual analog synthesis, Comput. Music J., vol. 30, no. 2, pp. 19–31,
2006, Summer.
[19] D. Lowenfels, “Virtual analog synthesis with a time-varying comb
filter, in Proc. Audio Eng. Soc. 115th Conv., New York, Oct. 2003,
paper no. 5960.
[20] J. Nam, V. Välimäki, J. S. Abel, and J. O. Smith, “Efficient imple-
mentation of anti-aliasing oscillators using low-order fractional delay
filters, IEEE Trans. Audio, Speech, Lang. Process., 2009, submitted
for publication.
[21] E. Terhardt, G. Stoll, and M. Seewann, “Algorithm for extraction of
pitch and pitch salience from complex tonal signals, J. Acoust. Soc.
Amer., vol. 71, no. 3, pp. 679–688, Mar. 1982.
[22] M. Bosi and R. Goldberg, Introduction to Digital Audio Coding and
. Norwell, MA: Kluwer, 2003.
Vesa Välimäki (S’90–M’92–SM’99) received the
M.Sc., the Licentiate of Science, and the Doctor
of Science degrees in technology, all in electrical
engineering, from the Helsinki University of Tech-
nology (TKK), Espoo, Finland, in 1992, 1994, and
1995, respectively. His doctoral dissertation dealt
with fractional delay filters and physical modeling
of musical instruments.
He was a Postdoctoral Research Fellow at the
University of Westminster, London, U.K., in 1996. In
1997 to 2001, he was Senior Assistant (cf. Assistant
Professor) at the TKK Laboratory of Acoustics and Audio Signal Processing,
Espoo, Finland. From 1998 to 2001, he was on leave as a Postdoctoral Re-
searcher under a grant from the Academy of Finland. In 2001 to 2002, he was
Professor of signal processing at the Pori unit of the Tampere University of
Technology, Pori, Finland. Since 2002, he has been Professor of audio signal
processing at TKK. He was appointed Docent (cf. Adjunct Professor) in signal
processing at the Pori unit of the Tampere University of Technology in 2003. In
2006–2007, he was the Head of the TKK Laboratory of Acoustics and Audio
Signal Processing. In 2008 to 2009, he was on sabbatical and spent several
months as a Visiting Scholar at the Center for Computer Research in Music and
Acoustics (CCRMA), Stanford University, Stanford, CA. His research interests
include sound synthesis, digital filters, and acoustics of musical instruments.
Prof. Välimäki is a member of the Audio Engineering Society, the Acoustical
Society of Finland, and the Finnish Musicological Society. He was President
of the Finnish Musicological Society from 2003 to 2005. In 2004, he was a
Guest Editor of the special issue of the
EURASIP Journal on Applied Signal
Processing on model-based sound synthesis. In 2008, he was the chairman of
DAFx-08, the 11th International Conference on Digital Audio Effects (Espoo,
Finland). In 2000–2001, he was the secretary of the IEEE Finland Section. From
2005 to 2009, he was an Associate Editor of the IEEE S
LETTERS. In 2007, he was a Guest Editor of the special issue of the IEEE Signal
Processing Magazine on signal processing for sound synthesis. He is currently
an Associate Editor of the IEEE T
PROCESSING and of the Research Letters in Signal Processing.Heis
a member of the Audio and Electroacoustics Technical Committee of the IEEE
Signal Processing Society. He is the Lead Guest Editor of the special issue of
the IEEE T
virtual analog audio effects and musical instruments.
Juhan Nam (S’09) was born in Busan, South Korea,
in 1976. He received the B.S. degree in electrical
engineering from Seoul National University, Seoul,
Korea, in 1998. He is currently pursuing the M.S.
degree in electrical engineering and the Ph.D. degree
at the Center for Computer Research in Music and
Acoustics (CCRMA), Stanford University, studying
signal processing applied to audio and music appli-
He was with Young Chang (Kurzweil Music
Systems) as a Software Engineer from 2001 to 2006,
working on software and DSP algorithms for musical keyboards.
Julius O. Smith received the B.S.E.E. degree in con-
trol, circuits, and communication from Rice Univer-
sity, Houston, TX, in 1975 and the M.S. and Ph.D.
degrees in electrical engineering from Stanford Uni-
versity, Stanford, CA, in 1978 and 1983, respectively.
His Ph.D. research was devoted to improved methods
for digital filter design and system identification ap-
plied to music and audio systems.
From 1975 to 1977, he worked in the Signal Pro-
cessing Department at ESL, Sunnyvale, CA, on sys-
tems for digital communications. From 1982 to 1986,
he was with the Adaptive Systems Department at Systems Control Technology,
Palo Alto, CA, where he worked in the areas of adaptive filtering and spectral es-
timation. From 1986 to 1991, he was with NeXT Computer, Inc., responsible for
sound, music, and signal processing software for the NeXT computer worksta-
tion. After NeXT, he became an Associate Professor at the Center for Computer
Research in Music and Acoustics (CCRMA) at Stanford University, teaching
courses and pursuing research related to signal processing techniques applied
to music and audio systems. Continuing this work, he is currently a Professor of
Music and Associate Professor of Electrical Engineering (by courtesy) at Stan-
ford University.
Jonathan S. Abel received the S.B. degree in
electrical engineering from the Massachusetts Insti-
tute of Technology, Cambridge, in 1982, where he
studied device physics and signal processing, and
the M.S. and Ph.D. degrees in electrical engineering
from Stanford University, Stanford, CA, in 1984
and 1989, respectively, focusing his research efforts
on statistical signal processing with applications to
passive sonar and GPS.
He is currently a Consulting Professor at the
Center for Computer Research in Music and Acous-
tics (CCRMA), Stanford University, where his research interests include audio
and music applications of signal and array processing, parameter estimation,
and acoustics. From 1999 to 2007, he was a Co-Founder and Chief Technology
Officer of the Grammy Award-winning Universal Audio, Inc. He was a Re-
searcher at NASA/Ames Research Center, exploring topics in room acoustics
and spatial hearing on a grant through the San Jose State University Founda-
tion. He was also chief scientist of Crystal River Engineering, Inc., where he
developed their positional audio technology, and a Lecturer in the Department
of Electrical Engineering at Yale University. As an industry consultant, he
has worked with Apple, FDNY, LSI Logic, NRL, SAIC, and Sennheiser, on
projects in professional audio, GPS, medical imaging, passive sonar, and fire
department resource allocation.
Prof. Abel is a Fellow of the Audio Engineering Society. He won the 1982
Ernst A. Guillemin Thesis Award for the best undergraduate thesis in the De-
partment of Electrical Engineering and Computer Science.
... Among these we can surely consider one family of methods descending from the Bandlimited Impulse Train (BLIT) [12], namely the bandlimited step (BLEP/minBLEP) [13] and polyBLEP [14]. Another family of methods stems from a different approach: these methods are known as Differentiated Polynomial Waveform (DPW) [15], and include the Polynomial Transition Regions (PTR) [16] and its more efficient form (EPTR) [17]. Additional approaches for generation of bandlimited waveforms are covered in [14]. ...
... Hence, an algorithm that generates a discrete-time parabolic waveform and differentiates it results in a sawtooth waveform with reduced aliasing since the spectral tilt of the parabolic waveform is steeper than that of a sawtooth. This approach can be extended to arbitrary order [15]: a polynomial waveform of order N is generated and then it is differentiated N − 1 times. In practice the order is bounded by finite precision arithmetic as each differentiation step is affected by quantization noise. ...
... One of the methods for sawtooth generation that has been widely discussed in the literature is the DPW method, which we have shown to be identical to AA-FIR-1. Anyway, we implemented DPW with polynomials of order 2 and 3 [15] (DPW-2 and DPW-3) to verify that they achieve exactly the same results as AA-FIR-1 and AA-FIR-2. ...
Full-text available
In the last decades many efficient methods have been proposed to generate waveforms with reduced aliasing, proving this field quite mature. However, the introduction of Antiderivative Antialiasing (AA) methods for the reduction of aliasing in nonlinear discrete-time processing, can shed a new light on bandlimited oscillators and provide a general method for dealing with aliasing in arbitrary waveforms generation. In this work we will first bridge the gap between AA methods and bandlimited waveform generation, and then show the flexibility of the recently introduced AA-IIR method for dealing with both classical synthesizer waveforms and arbitrary wavetables. We will show how AA methods can be used for periodic and non-periodic waveform generation, and provide an innovative AA-IIR method able to compute alias-reduced version of any waveform with arbitrary antialiasing filter order, without the effort of computing ad hoc analytical expressions. Antialiasing performance is compared with other well known methods, showing the effectiveness of the approach.
... Previous research on VA modeling of synthesizer circuits has concentrated on VCFs [6,7,8,9,10], oscillators [11,12,13], and effects processors [4,14,15]. Of related interest to this study is the pioneering work done during the 1970s on digital waveshaping synthesis [16,17,18]. ...
... 8-times oversampling w.r.t. standard audio rate) [13]. This sample rate is used throughout the rest of this study. ...
Conference Paper
Full-text available
The Serge Triple Waveshaper (TWS) is a synthesizer module de- signed in 1973 by Serge Tcherepnin, founder of Serge Modular Music Systems. It contains three identical waveshaping circuits that can be used to convert sawtooth waveforms into sine waves. However, its sonic capabilities extend well beyond this particular application. Each processing section in the Serge TWS is built around what is known as a Norton amplifier. These devices, unlike traditional operational amplifiers, operate on a current differencing principle and are featured in a handful of iconic musical circuits. This work provides an overview of Norton amplifiers within the context of virtual analog modeling and presents a digital model of the Serge TWS based on an analysis of the original circuit. Results obtained show the proposed model closely emulates the salient features of the original device and can be used to generate the complex waveforms that characterize “West Coast” synthesis.
... Indeed, it is well-known that geometric waveforms (sawtooth, square, triangle) are not bandlimited, therefore a digital oscillator that simply samples these waveforms yields aliasing distortion. In order to mitigate this problem, several oscillator algorithms have been proposed in the last decades, including those based on wavetable synthesis [1,2,3], oversampling [1], discrete summation formulas [4,5], frequency-domain methods [6,7], bandlimited impulse trains (BLITs) [8,9], bandlimited step functions (BLEPs) [10,11],differentiated polynomial waveforms (DPWs) [12,13], and polynomial transition regions (PTRs) [14,15]. The choice of the most appropriate algorithm depends on the efficiency and aliasing reduction requirements of the application at hand. ...
Conference Paper
Full-text available
Hard sync is a feature appearing in many analog synthesizers: it consists in retriggering a slave oscillator, regardless of its phase, every time a master oscillator completes its cycle. If this process is naïvely implemented digitally, it is subject to aliasing. While for sawtooth, square, and triangle waves several effective antialiasing methods have been developed, the literature is sparser concerning sine hard sync, arguably because discontinuities of infinite order are introduced which are more difficult to handle. In this paper, we introduce a new antialiasing algorithm for sine hard sync which is obtained by filtering the hard-synced sine with a FIR lowpass kernel, as opposed to existing methods based on the windowed sinc function. We show that our method yields lower computational cost and better aliasing reduction.
... The culmination was a simplistic modeling of the well-known Minimoog synthesizer. We built a band-limited sawtooth oscillator using the differentiated parabolic-wave algorithm (Valimaki et al. 2009), as well as a model of one stage of the ladder filter. This stage includes a first-order filter and soft saturation. ...
Full-text available
This article presents a method for programming musical signal-processing circuits visually, using expressive idioms and abstractions from functional programming. Special attention is paid to the creative workflow, framing the education in a constructionist context. Our aim is to empower musicians in signal processing: The claim was tested in a university workshop for relatively inexperienced programmers. The participants were able to study and implement signal-processing algorithms from literature and integrate them into their preexisting workflow, and appeared to gain self-confidence while doing so.
... Non-linear oscillators find widespread applications in science and engineering. Examples are found in the design and implementation of a liquid mixing apparatus based on a stirring tank model [56], in the design of digital music synthesizers [47], in the design of haptic devices [55] and in secure transmission of information over digital networks by signal masking [12], where synchronization of non-linear oscillator is also a topic of prime interest. ...
Full-text available
The present paper outlines a general second-order dynamical system on manifolds and Lie groups that leads to defining a number of abstract non-linear oscillators. In particular, a number of classical non-linear oscillators, such as the simple pendulum model, the van der Pol circuital model and the Duffing oscillator class are recalled from the dedicated literature and are extended to evolve on manifold-type state spaces. Also, this document outlines numerical techniques to implement these systems on a computing platform, derived from classical numerical schemes such as the Euler method and the Runke-Kutta class of methods, and illustrates their numerical behavior by a great deal of numerical examples and simulations.
... Up to now, the literature has dealt mainly with two issues in virtual analog oscillators: aliasing in the generation of geometrical waveforms (e.g., sawtooth or square), as well as analysis and emulation of specific circuits. Since the inception of virtual analog, several techniques have established to generate geometrical waveforms, namely BLIT [4], BLEP [8] (and variations thereof), BLAMP [9], DPW [10] and wavetable synthesis. Other studies addressed the peculiar behavior of analog circuits and their departure from the ideal behavior. ...
Conference Paper
Full-text available
Early analog synthesizer designs are very popular nowadays, and the discrete-time emulation of voltage-controlled oscillator (VCO) circuits is covered by a large number of virtual analog (VA) textbooks, papers and tutorials. One of the issues of well-known VCOs is their tuning instability and sensitivity to environmental conditions. For this reason, digitally-controlled oscillators were later introduce to provide stable tuning. Up to now, such designs have gained much less attention in the music processing literature. In this paper, we examine one of such designs, which is based on the Walsh-Hadamard transform. The concept was employed in the ARP Pro Soloist and in the Welson Syntex, among others. Some historical background is provided, along with a discussion on the principle, the actual implementation and a band-limited virtual analog derivation.
... A sizable portion of VA research has concentrated on designing efficient algorithms to generate alias-free geometric waveforms like those used in analog subtractive synthesizers, the so-called classic analog waveforms. Well-known techniques include the bandlimited impulse train (BLIT) family of methods, which involves the use of bandlimited basis functions and their integrated forms [39][40][41], and the use of differentiated polynomial waveforms (DPW) [42][43][44]. Moreover, Välimäki and Franck have applied the antialiasing principle behind the DPW algorithm to tackle aliasing in wavetable oscillators [45]. ...
Full-text available
Wavefolders are a particular class of nonlinear waveshaping circuits, and a staple of the “West Coast” tradition of analog sound synthesis. In this paper, we present analyses of two popular wavefolding circuits—the Lockhart and Serge wavefolders—and show that they achieve a very similar audio effect. We digitally model the input–output relationship of both circuits using the Lambert-W function, and examine their time- and frequency-domain behavior. To ameliorate the issue of aliasing distortion introduced by the nonlinear nature of wavefolding, we propose the use of the first-order antiderivative method. This method allows us to implement the proposed digital models in real-time without having to resort to high oversampling factors. The practical synthesis usage of both circuits is discussed by considering the case of multiple wavefolder stages arranged in series.
... These methods were first proposed in a signal processing context in [9], and recently extended to higher order in the distinct framework of numerical integration * S. Bilbao methods in [10], and have been shown to be competitive with oversampling methods, and allow for antialiasing, even for strong nonlinearities such as the hard clipper, at reduced sampling rates (the audio rate, or perhaps twice oversampled rate). Such antiderivativebased antialiasing methods bear some resemblance to methods used in the generation of audio waveforms with reduced aliasing (such as the "integrated wavetable" [11], and "differentiated parabolic waveform" or DPW [12,13]). One particular family of such antialiasing methods was proposed in [10]; these methods have the feature that, for nonlinearities which reduce to linear at low amplitudes, a low-pass filtering effect is present, increasing in strength with the order of antidifferentiation, and reducing the utility of such methods. ...
Conference Paper
Full-text available
Aliasing is major problem in any audio signal processing chain involving nonlinearity. The usual approach to antialiasing involves operation at an oversampled rate—usually 4 to 8 times an audio sample rate. Recently, a new approach to antialiasing in the case of memoryless nonlinearities has been proposed, which relies on operations over the antiderivative of the nonlinear function, and which allows for antialiasing at audio or near-audio rates, and without regard to the particular form of the nonlinearity (i.e., polynomial, or hard clipping). Such techniques may be deduced through an application of Lagrange interpolation over unequally-spaced values, and, furthermore, may be constrained to behave as spectrally transparent “throughs” for nonlinearities which reduce to linear at low signal amplitudes. Numerical results are presented.
The Rhodes piano is an electromechanical keyboard instrument, released for the first time in 1946 and subsequently manufactured for at least four decades, reaching an iconic status and being now generally referred to as the electric piano. A few academic works discuss its operating principle and propose different physical modeling strategies; however, the inharmonic modes that characterize the attack transient have not been subject of a dedicated study before. This study addresses this topic by first observing the spectrum at the pickup output, applying a psychoacoustic model to assess perceptual relevance, and then conducts a series of scanning laser Doppler vibrometry (SLDV) experiments on the Rhodes asymmetric tuning fork. This study compares the modes of the Rhodes piano to those of its individual parts, allowing for the extraction of important information regarding role and natural modes. On the basis of this study, numerical experiments are conducted that show the intermodulation of the modes due to the magnetic pickup and allow the tones produced by the Rhodes from the collected data to be closely matched. Finally, this study is able to extract the distribution of the most important modes found on the whole keyboard range of a Rhodes piano, which can be useful for sound synthesis.
Full-text available
Full-text available
This is the first book to develop both the theory and the practice of synthesizing musical sounds using computers. Each chapter starts with a theoretical description of one technique or problem area and ends with a series of working examples (over 100 in all), covering a wide range of applications. A unifying approach is taken throughout; chapter two, for example, treats both sampling and wavetable synthesis as special cases of one underlying technique. Although the theory is presented quantitatively, the mathematics used goes no further than trigonometry and complex numbers. The examples and supported software - along with a machine-readable version of the text - are available on the web and maintained by a large online community. The Theory and Techniques of Electronic Music is valuable both as a textbook and as professional reading for electronic musicians and computer music researchers. © 2007 by World Scientific Publishing Co. Pte. Ltd. All rights reserved.
Full-text available
A fast method is described for generating a periodic discrete‐time signal with harmonics of equal amplitude and a fundamental frequency which is not necessarily an integral fraction of the sampling frequency (as with ordinary pulse generators). Such a signal can be used as input to digital filters for the synthesis of speech and music.
Conference Paper
The bandlimited digital synthesis model of Stilson and Smith is extended with a single feed-forward comb filter. Time-varying comb filter techniques are shown to produce a variety of classic analog waveform effects, including waveform morphing, pulse-width modulation, harmonization, and frequency modulation. The techniques discussed are not guaranteed to maintain perfect bandlimiting; however, they are generally applicable to other synthesis models.
Foreword. Preface. I: Audio Coding Methods. 2. Quantization. 3. Representation of Audio Signals. 4. Time to Frequency Mapping Part I: The PQMF. 5. Time to Frequency Mapping Part II: The MDCT. 6. Introduction to Psychoacoustics. 7. Psychoacoustic Models for Audio Coding. 8. Bit Allocation Strategies. 9. Building a Perceptual Audio Decoder. 10. Quality Measurement of Perceptual Audio Codecs. II: Audio Coding Standards. 11. MPEG-1 Audio. 12. MPEG-2 Audio. 13. MPEG-2 AAC. 14.Dolby AC-3. 15. MPEG-4 Audio. Index.
This is a general introduction to the theory of computer music, giving details on sound, digital signal processing, math, and C programming. It assumes a strong knowledge of music.
A procedure is described for the automatic extraction of the various pitch percepts which may be simultaneously evoked by complex tonal stimuli. The procedure is based on the theory of virtual pitch, and in particular on the principle that the whole pitch percept is dependent both on analytic listening (yielding spectral pitch), and on holistic perception (yielding virtual pitch). The more or less ambiguous pitch percept governed by these two pitch modes is described by two pitch patterns: the spectral-pitch pattern, and the virtual-pitch pattern. Each of these patterns consists of a number of pitch (height) values, and associated weights, which account for the relative prominence of every individual pitch. The spectral-pitch pattern is constructed by spectral analysis, extraction of tonal components, evaluation of masking effects (masking and pitch shifts), and weighting according to the principle of spectral dominance. The virtual-pitch pattern is obtained from the spectral-pitch pattern by an advanced algorithm of subharmonic coincidence assessment. The procedure is based on a previous algorithm for calculating virtual pitch and can be regarded as a generalized and advanced version thereof.
BOOK: Prentice-Hall, 1990. 560pp. This is a general introduction to the theory of computer music, giving details on sound, digital signal processing, math, and the author's cmusic acoustic compiler. It assumes a strong knowledge of music.