David Bessell

Interdisciplinary Centre for Computer

Dynamic Convolution
Music Research (ICCMR)
Plymouth University Modeling, a Hybrid
Drake Circus, Plymouth
Devon PL4 8AA, UK Synthesis Strategy
David.Bessell@Plymouth.ac.uk
https://sites.google.com/site/davebessellmusic/home

Abstract: This article outlines a hybrid approach to the synthesis of percussion sounds. The synthesis method described
here combines techniques and concepts from physical modeling and convolution to produce audio synthesis of
percussive instruments. This synthesis method not only achieves a high degree of realism in comparison with audio
samples but also retains some of the flexibility associated with waveguide physical models. When the results are
analyzed, the method exhibits some interesting detailed spectral features that have some aspects in common with
the behavior of acoustic percussion instruments. In addition to outlining the synthesis process, the article discusses
some of the more creative possibilities inherent in this approach, e.g., the use and free combination of excitation and
resonance sources from beyond the realms of the purely percussive examples given.

Physical modeling offers the contemporary com-
poser the promise of many interesting and desirable
technical and creative attributes. Conceptually it
straddles the worlds of traditional acoustic and elec-
tronic composition, seeming to offer possibilities
to both. (As just one example, consider the virtual
creation of hybrid instruments made from disparate
components derived from real-world models but im-
possible to actually build in practice.) It is perhaps
surprising, then, that physical modeling has not
become more widely used in contemporary musical
works to date. Might there be some barrier that is
limiting the use of this technology by the wider
body of composers and instrumentalists?
To fully explore the potential offered by physical
modeling requires the acquisition of quite high-level
skills in computing, mathematics, and mechanical
and electrical engineering. Allowing the composer
to simulate conventional acoustic instruments—or
to create new instruments—that are expressive,
and that have specific timbral characteristics suited
to an individual piece, raises the question: At
what technical level must a musician engage to
fully exploit the possibilities? It is desirable for
a composer or player to have as much control
over the model as possible, to be truly able to
fit the instrument to the musical task at hand.
The acquisition of traditional musical skills is
no trivial task in itself, however; if we add to
Computer Music Journal, 37:1, pp. 44–51, Spring 2013
doi:10.1162/COMJ a 00159

c 2013 Massachusetts Institute of Technology.
this the demands of computing, mathematics,
and associated fields, there comes a point even
for the most gifted where the time involved and
learning curve become problematic. This has led to
a noticeable division between technically oriented
researchers and software developers, on the one
hand, and practicing musicians and composers, on
the other. The practical outcome of this is that
important advances in physical modeling research
are generally not being utilized creatively in the
majority of new music pieces.
Currently there are a variety of levels at which
a player or composer can engage with physical-
modeling technologies. There are (1) the pre-
built commercial instruments with fixed inter-
face options, such as Applied Acoustics’ String
Studio (Applied Acoustics 2011); (2) the modu-
lar systems with more or less pre-built building
blocks, exemplified by Native Instruments’ Reak-
tor (Native Instruments 2011) and by Modalys
(http://forumnet.ircam.fr/701.html?&L=1) from the
Institut de Recherche et Coordination Acoustique/
Musique (IRCAM); (3) graphical programming
environments, such as Cycling ‘74’s Max/MSP
(http://cycling74.com); (4) libraries for various text-
based software languages, such as STK, the Synthesis
ToolKit in C++ (Cook and Scavone 2004); and fi-
nally, (5) work from first principles implemented as
mathematical expressions in text-based code, such
as the work of Stefan Bilbao (Bilbao 2009). Each of
these approaches falls somewhere along a contin-
uum between maximum ease of use and maximum
flexibility. The approach proposed here, a variant

44 Computer Music Journal

on physical-modeling and convolution-synthesis
methods, offers a further option to this list. The
work focuses on percussive sounds in which an
attack part (source) and a decay part (filter) are
obtained from recorded sounds. The amplitude of
the attack component is used to derive an envelope
for a noise source.
snare-strike sample and 152 kB for the convolution
impulse sample. This potential for data reduction
in approaches based on physical modeling has been
noted by Rauhala, Lehtonen, and Välimäki (2007).
Limitations of Existing Techniques Used
in Isolation

Although each of the three fields mentioned earlier

A Hybrid Software Approach
has its own well-established body of techniques,
each also has its own strengths and limitations. For
Dynamic convolution modeling (DCM) is an ap-
example, convolution as normally implemented has
proach to realistic sound synthesis combining
a tendency to produce rather static aural results.
techniques and paradigms derived from the fields
Sampling likewise can only produce static snapshots
of physical modeling (Välimäki et al. 2006), real-
of audio, which require further manipulation to
time convolution (Stockham 1966; Gardner 1995),
render them more expressive in a musical context.
and audio sampling (Rabiner and Gold 1975).
Physical modeling can be powerfully expressive, but
Audio examples of the process in action were re-
at the expense of some fairly high-order complexity
alized with an implementation in Max/MSP. The
both in control and implementation. As noted,
Max/MSP implementation can be downloaded from
this complexity has led to some resistance to
https://sites.google.com/site/davebessellmusic/
its use among non-specialist musicians; note,
software-maxmsp.
for example, the relative commercial failure of
Software development in this project is ongoing,
Yamaha’s VL1 physical-modeling synthesizer, which
and this Max/MSP iteration is currently designed pri-
was launched in 1994 and was based on research
marily to recreate realistic percussion instruments
by Julius Smith (Smith 2005; Stanford University
that can be played from a standard MIDI keyboard,
2011). Dynamic convolution synthesis, as it is
although it already has some capabilities beyond
presented here, attempts to marry some of the
simple replication of realistic percussion. The mo-
conceptual and practical ease of use found in
tivation for attempting to create a hybrid synthesis
typical audio sampling software with some of the
method was to try to combine the strengths of some
realism of convolution filtering and reverberation
current staples of computer audio manipulation into
software, without entirely losing the flexibility and
a new expressive tool that could be easily managed
expressiveness of physical modeling.
by non-expert users or more traditionally oriented
composers while still maintaining sophisticated au-
dio and performance flexibility. During research into
a practical implementation of this concept, it was Precedents
discovered that, in comparison with the large and
somewhat unwieldy sample resources typically used A number of previous studies have elements in
for realistic and expressive drum emulation in com- common with the methods outlined here. There
mercial libraries, comparable audio results could be are precedents for the use of physical modeling for
achieved with significant data reduction while re- drums and percussion, such as the 2-D waveguide
taining a quite modest demand on CPU load. In com- percussion models by Van Duyne and Smith (1993),
parison to a well-known and popular drum sample li- the 3-D waveguides in the work of Laird (2001),
brary that uses 1.6 MB of data-compressed samples to and the “physically informed” percussion-synthesis
recreate a velocity-sensitive snare drum, the method approach of Perry Cook (Cook 1997). A number of
proposed here uses just 4 kB (uncompressed) for the approaches based on source-filter decomposition

Bessell 45

have also been proposed (Sandler 1990; Laroche and
Meillier 1994). Further work on physically modeled
percussion encompasses both modal approaches
and finite difference models (Avanzini and Marogna
2010; Rabenstein, Koch, and Popp 2010; Bilbao
2012), and there are precedents for the use of noise
inputs in a physical-modeling context, such as
Karplus and Strong (1983). Demoucron (2008) and
Demoucron and Rasamimanana (2009) have pre-
sented a violin system that uses physical modeling
for the input combined with convolution. Early
work on convolution as a synthesis tool was done
by Curtis Roads (Roads 1993). Mackenzie, Kale, and
Cain (1995) investigated percussion synthesis using
a variety of filters excited by a noise source; in this
case, the result of the uniform noise excitation is
then amplitude-scaled after the filter stage. Simi-
larly, Karjalainen et al. (2002) investigated the use of
autoregression, moving average, linear prediction,
and frequency zooming in a similar context. Macon
et al. (1998) implemented percussion synthesis using
an all-pole model that emphasizes efficiency.
Component Parts
The hybrid synthesis approach proposed here can be
broken down into a number of component parts, as
outlined here.
Physical Modeling
The principal element that the proposed hybrid
strategy takes from physical modeling is the concept
of separating the excitation source from the rest of
the process. In particular, techniques of noise exci-
tation seen in waveguide model implementations
(Smith 2006) were the starting point.
the excitation wavetable signals [were] obtained
by inverse filtering (deconvolution) of the
recorded sound by the SDL response. For
practical synthesis, only the initial transient
part of the inverse-filtered excitation is used,
typically covering several tens of milliseconds
(Välimäki et al. 2006).
46
Figure 1. Structure of the
synthesis process.
The particular form it takes in the current article,
however, is a little different from that commonly
seen in waveguide models. In this case, a noise
source is amplitude-modulated by an envelope
derived directly from an initial short transient
audio sample. (See the excitation sample, excitation
envelope, and pink noise sections of Figure 1.) This
amplitude-modulated noise source is then used as
the excitation for an impulse response derived from
an audio sample of a resonant body. (See Figure 1,
drum sample.) In order to generate this excitation
envelope, the initial transient sample waveform is
first rectified. This process is performed in order
to create a less colored result when the amplitude-
scaled noise source is convolved with the drum
resonance impulse. The excitation sample might
typically be drum-beater strike noise generated by
the user; see the subsequent section on sampling
for further details. The architecture of the whole
process can be seen in Figure 1. This architecture
draws generally on ideas found in the work of
Välimäki and Smith but is implemented here in
a way specific to this application; more on the
implications of this later.
Convolution
Developments in partitioned convolution im-
plementations (Gardner 1995; Stockham 1996)
combined with increases in computing speed now
mean that complex filter responses such as those
Computer Music Journal
associated with resonant objects can realistically be
implemented as part of a real-time process. In the
Max/MSP implementation utilized here, an external
object written by Alex Harker (Harker and Trem-
blay 2012) is used. This implements in Max/MSP
a similar strategy for real-time convolution as that
outlined by Gardner (1995). The Harker external
takes advantage of Macintosh-specific code to give
an efficient realization for this application. (Com-
mercial PC alternatives such as SIR2 or Freeverb3
could be substituted in this context.)
An overview of the real-time convolution process
that the Harker external uses is as follows. The
sample to be used as the impulse response is
partitioned, and the first partition is convolved in
the time domain using a process similar to that used
by the Max/MSP object buffir∼ (a finite-impulse-
response, or FIR, filter). An FIR filter convolves an
input signal with an impulse response, and this
can be implemented in a process known as direct
convolution. Direct convolution performed in the
time domain has no inherent latency but comes
with the disadvantage of a high computational cost.
This means that long convolutions such as the
ones needed for the DCM process are not viable
in real time using just FIR filters. In this case,
however, only the short first partition uses direct
convolution. The audio output of this first partition
is therefore available with zero latency. Playback of
this first-partition audio allows time for the delay
inherent in the frequency-domain convolution of the
second partition to elapse, and the second-partition
audio output can then be appended to that of the
first partition. This in turn allows time for the
longer delay associated with the larger fast-Fourier-
transform block size of the third partition, and so
on, in a cascade of just-in-time calculations.
Sampling
The current software implementation takes as its
starting point an audio sample of the drum, or
indeed any resonant body that is to be synthesized.
This “reverberant body” sample is used as the
impulse response for the convolution stage. Then
a second transient “strike” sample is needed as a
source for the amplitude envelope used to shape
the noise excitation for the convolution stage. This
second sample is created by sampling a damped
strike on the resonant body used for the first
sample. Alternative strike samples derived from
impacts on other resonant bodies using a variety of
beater materials and attack characters can also be
substituted, for a wider range of creative results like
those normally associated with physical modeling.
As one of our primary aims was user-friendliness,
techniques such as deconvolution as a possible
means of deriving strike samples from a single
overall source sample were not used. Combining
strike samples deconvolved from the same sample
as that of the resonant body would also have the
potential to introduce unwanted phase-related
coloring into the final audio result and would have
an associated increase in CPU usage. The two
samples—the resonant impulse and the beater strike
sample to be used as excitation envelope—are all
that is required to create a sense of realistic dynamic
behaviors. In an informal test, a small sample of
expert listeners (five) were consulted to confirm
the subjective audio-quality equivalence between
the results of the DCM process and conventional
commercial sample libraries. Readers can judge for
themselves by listening to audio samples created
with this process at http://soundcloud.com/dave
-bessell/drum-synthesis. [Editor’s note: The sound
examples will also appear on the DVD accompanying
Vol. 37, No. 4 (Winter 2013).]
The Sum of the Parts
Although the techniques associated with sampling,
convolution, and physical modeling have many
well-understood aspects, the particular configu-
ration suggested in this article gives some useful
results when synthesizing resonant bodies such as
drums.
Features
Combining the elements outlined previously allows
the user control over the following aspects. Separate
Bessell 47
samples can be loaded for the resonant-body impulse
and the beater strike. Audio output is triggered by
MIDI note input, and the user can specify the level
of variability and liveliness at the spectral level.
This control of variability in consecutive strikes
can be achieved by manipulating various factors,
including the pitch of the strike-sample playback
and the balance between the shaped noise excitation
and a more conventional single-sample excitation
impulse. MIDI velocity response is implemented by
scaling the amplitude of the strike envelope and the
frequency cutoff of a one-pole low-pass filter on the
output of this envelope (see Figure 1).
Creative Flexibility
Interestingly, this architecture retains some of the
flexibility and creative possibilities of the type of
physical model with which it shares some tech-
niques. A variety of different beater samples can be
used to “strike” the drum impulse. This variety of
strike excitations is by no means as flexible as the
hybrid acoustic/synthesized system proposed in the
work of Roberto Aimi (Aimi 2007), or the some-
what similar Korg Wavedrum commercial product.
The system proposed here is primarily aimed at a
user group oriented towards sampling technology
(rather than the live percussion players of Aimi’s
study), and for this user group the flexibility is
extended. The strike excitations can also include
nonrealistic combinations such as the pizzicato
gong hybrid heard at http://soundcloud.com/dave
-bessell/pizzgong, in which a pizzicato sample is
used in place of a conventional percussion beater
as a noise envelope for a gong sample. In this case,
the pizzicato excitation sample is created by record-
ing a heavily damped pizzicato on a violin string.
The MIDI control on the particular iteration of
the software used to create this example is some-
what more elaborate than the examples presented
so far.
This kind of creative flexibility was one of the
primary design aims for this approach. The general
approach is similar to that outlined by Rauhala
et al. (2007), although the method of excitation is
somewhat different.
48
A digital filter model for the soundboard has
been designed based on recorded bridge impulse
responses of the harpsichord. The output of the
string models is injected into the soundboard
filter that imitates the reverberant nature of the
soundbox . . . (Rauhala et al. 2007).
The excitation sample could be manipulated in
ways that give results comparable to those associated
with varying the hardness of the material that is used
to excite the resonant body in waveguide models.
This is not implemented in this particular example,
though, as it was found that more-convincing results
were obtained by just using a sample of a different
beater material in the first place.
Variability of Spectral Detail
As a consequence of the noise excitation method
used, the audio output also exhibits some properties
normally not associated with conventional sam-
pling. For instance, each audio response, although
broadly similar to the others, is nevertheless unique
in its micro detail at the spectral level. For an ex-
ample, see Figures 2 and 3 and compare the spectral
detail of two consecutive gong strikes created with
the DCM software.
This low-level spectral variability allows the
possibility of using multiple simultaneous layers
of audio without the obtrusive phase-cancellation
artifacts associated with static audio samples. Sim-
ilarly, multiple consecutive repeats of a sound are
each unique, obviating the need for the “round
robin” sample strategies typically found in com-
mercial sample libraries and playback engines.
One desirable aspect of conventional sampling and
convolution applications that is maintained, how-
ever, is the sense of convincing audio realism, to
the extent that it is almost impossible in some
contexts to distinguish the result from an actual
recording of the original drum or resonant body.
Particularly good results were obtained for this
implementation with metallic percussion such
as cymbals or gongs. Accompanying audio ex-
amples (http://soundcloud.com/dave-bessell/drum
-synthesis) illustrate DCM synthesis of a variety of
Computer Music Journal
Figure 2. DCM gong
strike 1.
percussion sounds: snare, gong, cymbals, tom-tom,
orchestral bass drum, and drumsticks.
Velocity Response
A further consequence of the combination of noise
excitation with the velocity-sensitivity imple-
mentation (see MIDI velocity, low-pass filter, and
amplitude in Figure 1), which is used as the input for
the convolution resonance section, is that to a great
degree the static nature of conventional convolution
is not apparent in the final result. Providing that the
original drum sample used as the impulse response
is sampled from a loud (fff) drum hit, a high degree
of realism in the velocity response can be achieved.
There are some small caveats to this aspect: A re-
sponse that is “louder” than the original resonance
sample cannot be created, and at extremely low
volumes there is some small subjective deviation
from the response of the actual drum in the real
world. In most practical musical situations this is
likely to be too small to be significant, but it could
easily be rectified by velocity-switching to a second
set of samples for the low-velocity extreme.
This strategy of placing the velocity sensitivity
outside the convolution process is conceptually
analogous to procedures followed in commuted
waveguide models (Smith 2010) and, combined
with the noise excitation, conveys a convincing
sense of dynamic response in the otherwise static
convolution output.
Notwithstanding some minor novel aspects in
the implementation of the noise excitation, the
individual elements mentioned previously are
largely well-established techniques. The particular
configuration presented in this work, however,
exhibits some surprisingly realistic, expressive, and
flexible audio attributes along with a significant
reduction in data storage, for results that can exceed
the realism of a conventional static sample. Typical
CPU load figures for this Max/MSP implementation
on a 2011 MacBook Pro for a single drum are in
the range of 1.5 to 2 percent. In short, the whole
Bessell 49
Figure 3. DCM gong
strike 2.
is considerably more than the sum of the parts
in terms of achievable audio performance and
user-friendliness for non-specialist users.
Outcomes and Areas for Further Research
There is the potential for further research into
extending this technique to wind and string in-
struments, but already this implementation allows
the free mixing of any resonant-body response
with any percussive-excitation response. Crucially,
these resonant bodies and percussive excitations
can easily be created by the user, thus in theory
allowing greater flexibility than modular physical
modeling approaches that rely on a menu of pre-
created building blocks. The possible excitation
sources include just about anything that can strike,
pluck, or scrape, and it is possible to imagine a more
abstract approach beyond the conventional, such as
excitation by water droplets. The user interface can
be made conceptually easy to manage, even for those
50
without specialist knowledge, while maintaining
a high degree of audio realism when compared to
conventional sampling. Some of the creative aspects
of conventional physical modeling techniques are
maintained, such as the possibilities for creating
expressive new hybrid instruments.
References
Aimi, R. M. 2007. “Hybrid Percussion: Extending Physical
Instruments Using Sampled Acoustics.” PhD disserta-
tion, Media Arts and Sciences, Massachusetts Institute
of Technology.
Applied Acoustics. 2011. “String Studio Overview.”
http://www.applied-acoustics.com/stringstudio/
overview/. Accessed 8 January.
Avanzini, F., and R. Marogna. 2010. “A Modular Physically
Based Approach to the Sound Synthesis of Membrane
Percussion Instruments.” IEEE Transactions on Audio,
Speech, and Language Processing 18(4):891–902.
Bilbao, S. 2009. Numerical Sound Synthesis: Finite Dif-
ference Schemes and Simulation in Musical Acoustics.
Chichester, UK: John Wiley and Sons.
Computer Music Journal
Bilbao, S. 2012. “Time Domain Simulation and Sound Syn-
thesis for the Snare Drum.” Journal of the Acoustical
Society of America 131(1):914–925.
Cook, P. R. 1997. “Physically Informed Sonic Modeling
(PhISM): Synthesis of Percussive Sounds.” Computer
Music Journal 21(3):38–49.
Cook, P. R., and G. P. Scavone. 2004. “The Synthesis
ToolKit (STK) in C++.” In K. Greenbaum, ed. Audio
Anecdotes: A Cookbook of Audio Algorithms and
Techniques. Natick, Massachusetts: A.K. Peters, pp.
237–253.
Demoucron, M. 2008. “On the Control of Virtual Violins:
Physical Modelling and Control of Bowed String
Instruments.” PhD thesis, L’Université Pierre et Marie
Curie and the Royal Institute of Technology (KTH
Stockholm).
Demoucron, M., and N. Rasamimanana. 2009. “Score
Based Real-Time Performance with a Virtual Violin.”
In Proceedings of the 12th International Conference on
Digital Audio Effects (DAFx-09). ID 38.
Gardner, W. G. 1995. “Efficient Convolution without
InputOutput Delay.” Journal of the Audio Engineering
Society 43(3):127–136.
Harker, A., and P. A. Tremblay. 2012. “The HISSTools Im-
pulse Response Toolbox: Convolution for the Masses.”
In Proceedings of the International Computer Music
Conference, pp. 148–155.
Karjalainen, M., et al. 2002. “Frequency-Zooming ARMA
Modeling of Resonant and Reverberant Systems.”
Journal of the Audio Engineering Society 50(12):1012–
1029.
Karplus, K., and A. Strong. 1983. “Digital Synthesis of
Plucked-String and Drum Timbres.” Computer Music
Journal 7(2):43–55.
Laird, J. 2001. “The Physical Modelling of Drums Using
Digital Waveguides.” PhD dissertation, Electrical and
Electronic Engineering, Bristol University, UK.
Laroche, J., and J. L. Meillier. 1994. “Multichannel
Excitation/Filter Modeling of Percussive Sounds with
Application to the Piano.” IEEE Transactions on Speech
and Audio Processing 2(2):329–344.
Mackenzie, J., I. Kale, and G. D. Cain. 1995. “Applying Bal-
anced Model Truncation to Sound Analysis/Synthesis
Models.” In Proceedings of the International Computer
Music Conference, pp. 400–403.
Macon, M., et al. 1998. “Efficient Analysis/Synthesis
of Percussion Musical Instrument Sound Using an
All-Pole Model.” In Proceedings of the IEEE Interna-
tional Conference on Acoustics, Speech, and Signal
Processing, pp. 3589–3592.
Native Instruments. 2011. “Reaktor 5.”
http://www.native-instruments.com/#/en/products/
producer/reaktor-5/. Accessed 9 January 2012.
Rabenstein, R., T. Koch, and C. Popp. 2010. “Tubular Bells:
A Physical and Algorithmic Model.” IEEE Transactions
on Audio, Speech, and Language Processing 18(4):881–
890.
Rabiner, L. R., and B. Gold. 1975. Theory and Application
of Digital Signal Processing. Englewood Cliffs, New
Jersey: Prentice-Hall.
Rauhala, J., H. M. Lehtonen, and V. Välimäki. 2007. “To-
ward Next-Generation Digital Keyboard Instruments.”
IEEE Signal Processing Magazine 24(2):12–20.
Roads, C. 1993. “Musical Sound Transformation by
Convolution.” In Proceedings of the International
Computer Music Conference, pp. 102–109.
Sandler, M. 1990. “Analysis and Synthesis of Atonal
Percussion Using High Order Linear Predictive Coding.”
Applied Acoustics 30(2–3):247–264.
Smith, J. O. 2005. “Physical Audio Signal Processing
for Virtual Musical Instruments and Audio Effects.”
https://ccrma.stanford.edu/∼jos/pasp05/. Last modified
December 2005; accessed 12 December 2010.
Smith, J. O. 2006. “A Basic Introduction to Digital
Waveguide Synthesis (for the Technically Inclined).”
https://ccrma.stanford.edu/∼jos/swgt/swgt.html. Last
modified February 2006; accessed 13 December
2010.
Smith, J. O. 2010. “Commuted Waveguide Synthesis.”
https://ccrma.stanford.edu/∼jos/cs.html. Accessed 12
December 2010.
Stanford University. 2011. “Patents and Applications.”
www.sondiusxg.com/patent.html. Accessed 2 February
2011.
Stockham, T. G. 1966. “High-Speed Convolution and
Correlation.” In AFIPS Proceedings 1966 Spring Joint
Computer Conference, Vol. 28, pp. 229–233.
Välimäki, V., et al. 2006. “Discrete-Time Modelling of
Musical Instruments.” Reports on Progress in Physics
69(1):1–78.
Van Duyne, S. A., and J. O. Smith. 1993. “Physical
Modeling with the 2-D Digital Waveguide Mesh.”
In Proceedings of the International Computer Music
Conference, pp. 40–47.
Bessell 51