C´arthach´O Nuan´ain, Perfecto
Herrera, and Sergi Jord `a
Music Technology Group
Communications Campus–Poblenou
Universitat Pompeu Fabra
Carrer Roc Boronat, 138, 08018
Barcelona, España
{carthach.onuanain, perfecto.herrera,
sergi.jorda}@upf.edu
Rhythmic Concatenative
Synthesis for Electronic
Music: Techniques,
Implementation, y
Evaluation
Abstracto: In this article, we summarize recent research examining concatenative synthesis and its application and
relevance in the composition and production of styles of electronic dance music. We introduce the conceptual
underpinnings of concatenative synthesis and describe key works and systematic approaches in the literature. Nuestro
sistema, RhythmCAT, is proposed as a user-friendly system for generating rhythmic loops that model the timbre and
rhythm of an initial target loop. The architecture of the system is explained, and an extensive evaluation of the system’s
performance and user response is discussed based on our results.
Históricamente, reusing existing material for the pur-
poses of creating new works has been a widely
practiced technique in all branches of creative arts.
The manifestations of these expressions can be
wholly original and compelling, or they may be
derivative, uninspiring, and potentially infringe on
derechos de autor (depending on myriad factors including
the domain of the work, the scale of the reuse, y
cultural context).
In the visual arts, reusing or adapting existing ma-
terial is most immediately understood in the use of
collage, where existing works or parts thereof are as-
sembled to create new artworks. Cubist artists such
as Georges Braque and Pablo Picasso extensively ref-
erenced, appropriated, and reinterpreted their own
works and the works of others, as well as common
found objects from their surroundings (Greenberg
1971). Collage would later serve as direct inspiration
for bricolage, reflecting wider postmodernist trends
towards deconstructionism, self-referentiality, y
revisionism that include the practice of parody and
pastiche (Lochhead and Auner 2002).
In music and the sonic arts, the natural corollary
of collage came in the form of musique concr `ete
(holmes 2008), a movement of composition stem-
ming from the experiments of Pierre Schaeffer and,
más tarde, Pierre Henry at the studios of Radiodiffusion-
T ´el ´evision Franc¸ aise in Paris during the 1940s and
1950s (Battier 2007). In contrast to the artificially
Computer Music Journal, 41:2, páginas. 21–37, Verano 2017
doi:10.1162/COMJ a 00412
C(cid:2) 2017 Instituto de Tecnología de Massachusetts.
and electronically generated elektronische Musik
spearheaded by Karlheinz Stockhausen at the West
German Radio studios in Cologne, the French com-
posers sought to conceive their works from existing
recorded sound, including environmental sources
like trains and speech. Seemingly unrelated and
nonmusical sounds are organized in such a way that
the listener discovered the latent musical qualities
and structure they inherently carry.
It is important to note that in music composi-
tion general appropriation of work predates these
electronic advancements of technology. In West-
ern art music, Por ejemplo, composers like B ´ela
Bart ´ok—himself a musicologist—have often turned
to folk music for its melodies and dance music styles
(Bart ´ok 1993), y otros (p.ej., Claude Debussy, cf.
Tamagawa 1988) became enchanted by music from
other cultures, such as Javanese gamelan, studying
its form and incorporating the ideas into new pieces.
Quotations, or direct lifting of melodies from other
composers’ works, are commonplace rudiments in
jazz music. Charlie Parker, Por ejemplo, was known
to pepper his solos with reference to Stravinsky’s
Rite of Spring (Mangani, Baldizzone, and Nobile
2006). David Metzer has compiled a good reference
on appropriation and quotation music (Metzer 2003).
The modern notion of sampling stems from the
advent of the digital sampler and its eventual explo-
sion of adaptation in hip-hop and electronic music.
Artists such as Public Enemy and the Beastie Boys
painstakingly assembled bewildering permutations
of musical samples, sound bites, and other miscel-
laneous recorded materials that sought to supplant
´O Nuan ´ain et al.
21
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
the many cultural references that permeated their
lyrics (Sewell 2014). Más tarde, the influence of hip-hop
production would inform the sample-heavy arrange-
ments of jungle and drum and bass, En particular
with its exhaustive rerendering of the infamous
“Amen Break.” John Oswald, an artist who directly
challenged copyright for artistic gain, dubbed his ap-
proach “plunderphonics” and set out his intentions
in a suitably subtitled essay “Plunderphonics, or Au-
dio Piracy as a Compositional Prerogative” (Oswald
1985). Using tape-splicing techniques, he created
deliberately recognizable montages of pop music,
such as that by Michael Jackson, in a style that be-
came later known as “mashups.” Nowadays, artists
such as Girltalk create extremely complex and mul-
tireferential mashups of popular music, harnessing
the powerful beat-matching and synchronization
capabilities of the modern digital audio workstation
(Humphrey, Turnbull, and Collins 2013).
Although the question of originality and author-
ship is not in the realm of this discussion, este
interesting and pertinent topic is under the scrutiny
of researchers in musicology and critical studies.
We encourage the reader to consult work by Tara
Rodgers (2003), Paul Miller (2008), and Kembrew
McLeod (2009) for a more focused discourse.
Associated research efforts in computer music,
signal processing, and music information retrieval
(MIR) afford us the opportunity to develop au-
tomated and intelligent systems that apply the
aesthetic of sampling and artistic reuse. The term
concatenative synthesis has been extensively used
to describe musical systems that create new sound
by automatically recycling existing sounds ac-
cording to some well-defined set of criteria and
algorithmic procedures. Concatenative synthesis
can be considered the natural heir of granular syn-
tesis (Roads 2004), a widely examined approach
to sound synthesis using tiny snippets (“grains”)
of around 20–200 msec of sound, which traces its
history back to Iannis Xenakis’s theories in For-
malized Music (Xenakis 1971). With concatenative
synthesis, the grains become “units” and are more
related to musical scales of length, such as notes
and phrases. Más importante, information is at-
tached to these units of sound: crucial descriptors
that allow spectral and temporal characteristics
of the sound to determine the sequencing of final
producción.
In the following sections, we will present a
thorough, critical overview of many of the key
works in the area of concatenative synthesis, based
on our observation that there has not been such
a broad survey of the state of the art in other
publications in recent years. We will compare
and contrast characteristics, técnicas, y el
challenges of algorithmic design that repeatedly
arise. For the past three years, we have been working
on the European-led initiative GiantSteps (Knees et
Alabama. 2016). The broad goal of the project is the research
and development of expert agents for supporting and
assisting music makers, with a particular focus
on producers of electronic dance music (EDM).
Como consecuencia, one of the focuses of the project has
been on user analysis: thinking about their needs,
desires, and skills; investigating their processes
and mental representations of tasks and tools; y
evaluating their responses to prototypes.
Modern EDM production is characterized by
densely layered and complex arrangements of tracks
making liberal use of synthesis and sampling,
exploiting potentially unlimited capacity and pro-
cessing in modern computer audio systems. Uno
of our main lines of research in this context has
been the investigation of concatenative synthesis
for the purposes of assisting music producers to
generate rhythmic patterns by means of automatic
and intelligent sampling.
In this article, we present the RhythmCAT
sistema, a digital instrument that creates new
loops emulating the rhythmic pattern and timbral
qualities of a target loop using a separate corpus of
sound material. We first proposed the architecture
of the system in a paper for the conference on New
Interfaces for Musical Expression ( ´O Nuan ´ain, Jord `a,
and Herrera 2016a), followed by papers evaluating
it in terms of its algorithmic performance ( ´O
Nuan ´ain, Herrera, and Jord `a 2016) and a thematic
analysis of users’ experience ( ´O Nuan ´ain, Jord `a,
and Herrera 2016b). This article thus represents
an expanded synthesis of the existing literature,
our developments motivated by some detected
shortcomings, and the illustration of an evaluation
estrategia.
22
Computer Music Journal
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
State of the Art in Concatenative Synthesis
Other authors have previously provided insightful
summaries of research trends in concatenative
synthesis (p.ej., Schwarz 2005; Sturm 2006). Estos
surveys are over ten years old, sin embargo (but see
Schwarz 2017 for a continuously updated online
survey), so we offer here a more recent compendium
of state-of-the-art systems as we see them, Residencia en
our investigations of previous publications up until
now.
Before music, concatenative synthesis enjoyed
successful application in the area of speech syn-
tesis; Hunt and Black (1996) first reported a unit
selection scheme using hidden Markov models
(HMM) to automatically select speech phonemes
from a corpus and combine them into meaningful
and realistic sounding sentences. Hidden Markov
models extend Markov chains by assuming that
“hidden” states output visible symbols, y el
Viterbi algorithm (Rabiner 1989) can return the
most probable sequence of states given a particular
sequence of symbols. In concatenative synthesis,
the maximum probabilistic model is inverted to
facilitate minimal cost computations.
The target cost of finding the closest unit in
the corpus to the current target unit becomes the
emission probability, with the concatenation cost
representing the transition probability between
estados. The Viterbi algorithm thus outputs indices
of database units corresponding to the optimal
state sequence for the target, based on a linear
combination of the aforementioned costs. Diemo
Schwarz (2003) directly applied this approach for
musical purposes in his Caterpillar system.
Schwarz notes, sin embargo, that the HMM approach
can be quite rigid for musical purposes because it
produces one single optimized sequence without
the ability to manipulate the individual units.
To address these limitations, he reformulates
the task into a constraint-satisfaction problem,
which offers more flexibility for interaction. A
constraint-satisfaction problem models a problem
as a set of variables, valores, and a set of constraints
that allows us to identify which combinations
of variables and values are violations of those
constraints, thus allowing us to quickly reduce large
portions of the search space (Russell and Norvig
2009).
Zils and Pachet (2001) first introduced constraint
satisfaction for concatenative synthesis in what
they describe as musical mosaicking—or, to use
their portmanteau, musaicing. They define two
categories of constraints: segment and sequence
constraints. Segment constraints control aspects of
individual units (much like the target cost in an
HMM-like system) based on their descriptor values.
Sequence constraints apply globally and affect
aspects of time, continuity, and overall distributions
of units. The constraints can be applied manually
by the user or learned by modeling a target. El
musically tailored “adaptive search” algorithm
performs a heuristic search to minimize the total
global cost generated by the constraint problem.
One immediate advantage of this approach over the
HMM is the ability to run the algorithm several
times to generate alternative sequences, mientras
the Viterbi process always outputs the most optimal
solución.
A simpler approach is presented in MatConcat
(Sturm 2004), using feature vectors comprising six
descriptors and computing similarity metrics be-
tween target units and corpus units. Built for the
MATLAB environment for scientific computing,
the interface is quite involved, and the user has
control over minute features such as descriptor
tolerance ranges, relative descriptor weightings,
as well as window types and hop sizes of output
transformations. On Sturm’s Web site are short
compositions generated by the author using ex-
cerpts from a Mahler symphony as a target, y
resynthesized using various unrelated sound sets,
por ejemplo, pop vocals, found sounds, and solo in-
strumental recordings from saxophone and trumpet
(www.mat.ucsb.edu/∼b.sturm/music/CVM.htm).
As concatenative synthesis methods matured,
user modalities of interaction and control became
more elaborate and real-time operations were
introducido. One of the most compelling features
of many concatenative systems is the concept of
the interactive timbre space. With the release of
CataRT (Schwarz et al. 2006), these authors provided
an interface that arranges the units in an interactive
two-dimensional timbre space. The arrangement
´O Nuan ´ain et al.
23
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
of these units is according to a user-selectable
descriptor on each axis. Instead of using a target
sound file to inform the concatenation procedure,
the user’s mouse cursor becomes the target. Sounds
that are within a certain range of the mouse cursor
are sequenced according to some triggering options
(one-shot, loop, and—most crucially—with real-
time output).
Bernardes takes inspiration from CataRT and
from Tristan Jehan’s Skeleton (Jehan 2005) a
build his EarGram system for the Pure Data (Pd)
ambiente (Bernardes, Guedes, and Pennycook
2013). Built on top of William Brent’s excellent
feature-extraction library timbreID (Brent 2010), él
adds a host of interesting features for visualization
and classification. Por ejemplo, así como el
familiar waveform representation and previously
described 2-D timbre representation (with various
clustering modes and dimensionality-reduction
implementations), there are similarity matrices
that show the temporal relations in the corpus
con el tiempo. Some unique playback and sequencing
modes also exist, such as the infiniteMode, cual
generates endless playback of sequences, y el
soundscapeMap, which features an additional 2-D
control of parameters pertaining to sound scene
diseño. Another system that adapts a 2-D timbre
space is AudioGarden by Frisson, Picard, y
Tardieu (2010), which offers two unique mapping
procedures. The first of these, “disc” mode, lugares
units by assigning the length of the audio file to the
radius of the unit from the center, with the angle
of rotation corresponding to a principal component
of timbre, mel-frequency cepstrum coefficients
(MFCCs). In the other mode, called “flower” mode,
a point of the sound is positioned in the space
according to the average MFCCs of the entire sound
archivo. Segments of the particular sound are arranged
in chronological fashion around this center point.
There have been some concatenative systems tai-
lored specifically with rhythmic purposes in mind.
Pei Xiang proposed Granuloop for automatically re-
arranging segments of four different drum loops into
a 32-step sequence (Xiang 2002). Segmentation is
done manually, without the aid of an onset detector,
using the Recycle sample editor from Propellerhead
Software. Segmented sounds are compared using the
inner product of the normalized frequency spectrum,
supplemented with the weighted energy. These val-
ues become weights for a Markov-style probability
transition matrix. Implemented in Pd, the user
interacts by moving a joystick in a 2-D space, cual
affects the overall probability weightings determin-
ing which loop segments are chosen. El sistema
presents an interesting approach but is let down by
its lack of online analysis. Ringomatic (Aucouturier
and Pachet 2005) is a real-time agent specifically
tailored for combining drum tracks, expanding on
many of the constraint-based ideas from their prior
musaicing experiments. They applied the system to
real-time performance following symbolic feature
data extracted from a human MIDI keyboard player.
They cite, as an example, that a predominance of
lower-register notes in the keyboard performance
applies an inverse constraint that creates comple-
mentary contrast by specifying that high-frequency
heavy cymbal sounds should be concatenated.
As demonstrated in EarGram, concatenative
synthesis has been considered useful in sound
design tasks, allowing the sound designer to build
rich and complex textures and environments that
can be transformed in many ways, both temporally
and timbrally. Cardle, Arroyos, and Robinson (2003)
describe their Directed Sound Synthesis software as a
means of providing sound designers and multimedia
producers a method of automatically reusing and
synthesizing sound scenes in video. Users select
one or more regions of an existing audio track and
can draw probability curves on the timeline to
influence resynthesis of these regions elsewhere
(one curve per region). Hoskinson and Pai (2001), en
a nod to granular synthesis, refer to the segments
used in their Soundscapes software as ”natural
grains,” and they seek to synthesize endless streams
of soundscapes. The selection scheme by which
segments are chosen is based on a representation
of each segment as a transition state in a Markov
cadena. Its interface features knobs and sliders for
interactively controlling gain and parameters of
multiple samples. To evaluate the platform they
conducted an additional study (Hoskinson and Pai
2007) to reveal whether listening subjects found the
concatenated sequences convincing compared with
genuinely recorded soundscapes.
24
Computer Music Journal
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
Cifra 1. Block diagram of
functionality in the
RhythmCAT system.
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
Cifra 1 gives a diagrammatic overview of these
important stages, which can be briefly summarized
como:
More-specific and applied-use cases of concate-
native synthesis include work by Ben Hackbarth,
who explores the possibilities of concatenative syn-
thesis in large-scale music composition (Hackbarth,
Schnell, and Schwarz 2011). Hackbarth has worked
with Schwarz to provide an alternative interface
for exploring variations based on a force-directed
graph. John O’Connell describes a graphical system
for Pd that demonstrates the use of higher-level
perceptual concepts like mood (happy versus sad)
for informing selection in audio mosaics (O’Connell
2011).
Commercial implementations also exist for con-
catenative synthesis. Of particular note is Steinberg’s
Loopmash, a software plug-in and mobile application
for automatically creating mashups from existing
looped content (www.steinberg.net/loopmash). El
interface consists of a number of tracks in a timeline
arrangement. One track is set as a master, and slices
in the master are replaced with matching slices from
the other slave tracks. Users interact by manipulat-
ing “similarity gain” sliders that control the influ-
ence of each track in the slice selection algorithm.
Other applications exist more as MIDI sampler sys-
tems attempting to model the performance qualities
of natural sources such as orchestral ensembles (p.ej.,
SynfulOrchestra, www.synful.com) or the human
voice (p.ej., Vocaloid, www.vocaloid.com).
There are many other concatenative systems that
are too numerous to discuss in detail here. Tenemos,
sin embargo, compiled a table in a previous publication
summarizing all the systems we have come across
in our research, with remarks on interaction and
visualization features, support for rhythm, y
whether any user evaluation was carried out ( ´O
Nuan ´ain, Jord `a, and Herrera 2016b).
Design and Implementation
En esta sección, we will describe our implementa-
tion of the RhythmCAT system, beginning with
an explanation of the musical analysis stages of
onset detection, segmentation, and feature extrac-
ción. This is followed by an examination of the
interactive user interface and the pattern-generation
proceso.
1. Sound Input
2. Onset Detection and Segmentation
3. Audio Feature Extraction
4. Storage and Data Representation
5. Pattern Synthesis
6. Real-Time Audio Output
´O Nuan ´ain et al.
25
The system is developed in C++ using the
JUCE framework (www.juce.com), the Essentia
musical analysis library (Bogdanov et al. 2013), y
the OpenCV computer vision library for matrix
operaciones (Bradski 2000).
Sound Input
The first stage in building a concatenative music
system generally involves gathering a database of
sounds from which selections can be made during
the synthesis procedure. This database can be
manually assembled, but in many musical cases the
starting point is some user-provided audio that may
range in length from individual notes to phrases to
complete audio tracks.
The two inputs to the system are the sound
palette and the seed sound. The sound palette refers
to the pool of sound files we want to use as the
sample library for generating our new sounds. El
seed sound refers to the short loop that we wish
to use as the similarity target for generating those
sounds. The final output sound is a short (one to two
barras) loop of concatenated audio that is rendered in
real time to the audio host.
Onset Detection and Segmentation
In cases where the sounds destined for the sound
palette exceed note or unit length, the audio needs
to be split into its constituent units using onset
detection and segmentation.
Onset detection is a large topic of continuous
estudiar, and we would encourage the reader to exam-
ine the excellent review of methods summarized
by Simon Dixon (2006). Actualmente, with some tun-
ing of the parameters, Sebastien Bock’s Superflux
algorithm represents one of the best-performing
state-of-the-art detection methods (B ¨ock and Wid-
mer 2013). For our purposes, we have experienced
good results with the standard onset detector avail-
able in Essentia, which uses two methods based on
analyzing signal spectra from frame to frame (at a
rate of around 11 mseg). The first method involves
estimating the high-frequency content in each frame
(Masri and Bateman 1996) and the second method
involves estimating the differences of phase and
magnitude between each frame (Bello and Daudet
2005).
The onset detection process produces a list of
onset times for each audio file, which we use to
segment into new audio files corresponding to unit
sounds for our concatenative database.
Audio Feature Extraction
In MIR systems, the task of deciding which features
are used to represent musical and acoustic properties
is a crucial one. It is a trade-off between choosing
the richest set of features capable of succinctly
describing the signal, Por un lado, y el
expense of storage and computational complexity,
en el otro. When dealing specifically with musical
signals, there are a number of standard features cor-
responding roughly to certain perceptual sensations.
We briefly describe the features we chose here (for a
more thorough treatment of feature selection with
relation to percussion, see Herrera, Dehamel, y
Gouyon 2003; Tindale et al. 2004; and Roy, Pachet,
and Krakowski 2007).
Our first feature is the loudness of the signal,
which is implemented in Essentia according to
Steven’s Power Law, a saber, the energy of the
signal raised to the power of 0.67 (Bogdanov et al.
2013). This is purported to be a more perceptually
effective measure for human ears. Próximo, we extract
the spectral centroid, which is defined as the
weighted mean of the spectral bins extracted using
the Fourier transform. Each bin is then weighted by
its magnitude.
Perceptually speaking, the spectral centroid
relates mostly to the impression of the brightness
of a signal. In terms of percussive sounds, uno
would expect the energy of a kick drum to be more
concentrated in the lower end of the spectrum and
hence have a lower centroid than that from a snare
or crash cymbal.
Another useful single-valued spectral feature is
the spectral flatness. It is defined as the geomet-
ric mean of the spectrum divided by the arithmetic
mean of the spectrum. A spectral flatness value of 1.0
26
Computer Music Journal
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
means the energy spectrum is flat, whereas a value of
0.0 would suggest spikes in the spectrum indicating
harmonic tones (with a specific frequency). El
value intuitively implies a discrimination between
noisy or inharmonic signals and signals that are
harmonic or more tonal. Kick-drum sounds (especialmente-
cially those generated electronically) often comprise
quite a discernible center frequency, whereas snares
and cymbals are increasingly broadband in spectral
energía.
Our final feature is MFCCs. These can be con-
sidered as a compact approximation of the spectral
envelope and is a useful aid in computationally
describing and classifying the timbre of a signal. Él
has been applied extensively in speech processing,
genre detection (Tzanetakis, Essl, and Cook 2001),
and instrument identification (Loughran et al. 2004).
The computation of MFCCs, as outlined by Beth
logan (2000), is basically achieved by computing
the spectrum, mapping the result into the more
perceptually relevant mel scale, taking the log, y
then applying the discrete cosine transform.
It is difficult to interpret exactly what each of the
MFCC components mean, but the first component
is generally regarded as encapsulating the energy.
Because we are already extracting the loudness
using another measure, we have discarded this
component in our system. For detailed explanations
and formulae pertaining to the features introduced
aquí, as well as others, we direct the reader to
Geoffroy Peeters’s compendium (Peeters 2004).
Storage and Data Representation
Further on in this article we will describe in greater
detail how the seed or target audio signal is actually
received from the Virtual Studio Technology host,
but in terms of analysis on that seed signal, el
process is the same as before: onset detection and
segmentation followed by feature extraction.
The resulting feature vectors are stored in two
matrices: the palette matrix and the target matrix.
The palette matrix stores the feature vectors of each
unit of sound extracted from the sound palette, y
the target matrix similarly stores feature vectors of
units of sound extracted from the seed loop.
Pattern Synthesis and Real-Time Audio Output
This section details the visible, aural, and interactive
elements of the system as they pertain to the user.
Cifra 2 provides a glimpse of the user interface in a
typical pattern generation scenario.
Workflow
The layout of the interface was the result of a
number of iterations of testing with users who,
while praising the novelty and sonic value of
the instrument, sometimes expressed difficulty
understanding the operation of the system. Uno de
the main challenges faced was how best to present
the general workflow to the user in a simple and
concise manner. We decided to represent the flow of
the various operations of the software emphatically
by using a simple set of icons and arrows, as seen in
Figure 2a.
The icons indicate the four main logical opera-
tions that the user is likely to implement, and opens
up related dialog screens:
Palette Dialog – indicated by the folder icon
Seed Dialog – indicated by the jack cable icon
Sonic Dialog – indicated by the square feature
space icon
Output Dialog – indicated by the speaker icon
Sound Palette
The user loads a selection of audio files or folders
containing audio files that are analyzed to create
the sound palette, as has previously been discussed.
Próximo, dimensionality reduction is performed on
each feature vector of the units in the sound
palette using principal component analysis (PCA).
Two PCA components are retained and scaled
to the visible area of the interface to serve as
coordinates for placing a circular representation
of the sound in two-dimensional space. We call
these visual representations, along with their
associated audio content, sound objects. Ellos son
clearly visible in the main Timbre Space window,
Figure 2d.
´O Nuan ´ain et al.
27
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
Cifra 2. The main user
interface for RhythmCat
consists of panels for
workflow (a), slider
controls (b), master
controls (C), el principal
timbre space interface (d),
and waveform
representación (mi).
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
Seed Input
Seed audio is captured and analyzed by directly
recording the input audio of the track on which
the instrument resides in the audio host. Usando
the real-time tempo and information about bar and
beat position provided by the host, the recorder
will wait until the next measure starts to begin
capture and will only capture complete measures
of audio. This audio is analyzed as before, with one
exception. Because the goal of the instrument is
to integrate with an existing session and generate
looped material, we assume that the incoming audio
is quantized and matches the tempo of the session.
De este modo, onset detection is not performed on the seed
aporte; en cambio, segmentation takes place at the
points in time determined by the grid size (más bajo
left of the screen).
An important aspect to note: Because the instru-
ment fundamentally operates in real time, nosotros necesitamos
to be careful about performing potentially time-
consuming operations, such as feature extraction,
when the audio system is running. De este modo, we perform
the audio-recording stage and feature-extraction pro-
cess on separate threads, so the main audio-playback
thread is uninterrupted. This is separate to yet
another thread that handles elements of the user
interface.
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
Sonic Parameters
Clicking on the square sonic icon in the center
of the workflow component opens up the set of
sliders shown in Figure 2b, which allows us to
adjust the weights of the features in the system.
Adjusting these weights has effects in terms of
the pattern-generation process but also in the
visualization. Presenting their technical names
(centroid, flatness, and MFCCs) would be confusing
28
Computer Music Journal
Cifra 3. Algorithm for
generating a list of sound
connections.
for the general user, so we relabeled them with
what we considered the most descriptive subjective
terms. With the pattern-generation process, estos
weights directly affect the features when performing
similarity computation and unit selection, as we
will see in the next section. Depending on the
source and target material, different combinations
of feature weightings produce noticeably different
resultados. Informalmente, we have experienced good
results using MFCCs alone, Por ejemplo, así como
combinations of the flatness and centroid. In terms
of visualization, when the weights are changed,
dimensionality reduction is reinitiated and, hence,
positioning of the sound objects in the timbre space
cambios. Manipulating these parameters can help
disperse and rearrange the sound objects for clearer
interaction and exploration by the user in addition
to affecting the pattern generation process.
Once the palette and seed matrices have been
populated, a similarity matrix between the palette
and seed matrix is created. Using the feature
weightings from the parameter sliders, a sorted
matrix of weighted Euclidean distances between
each onset in the target matrix and each unit sound
in the palette matrix is computed.
Unit Selection and Pattern Generation
The algorithm for unit selection is quite straight-
forward. For each unit i in the segmented target
secuencia (p.ej., a 16-step sequence) and each corpus
unit j (typically many more), the target unit cost
Ci, j is calculated by the weighted Euclidean distance
of each feature k.
These unit costs are stored in similarity matrix
METRO. Next we create a matrix M(cid:3) of the indices
of the elements of M sorted in ascending order.
Finalmente, a concatenated sequence can be generated
by returning a vector of indices I from this sorted
matrix and playing back the associated sound file.
To retrieve the closest sequence V0 one would only
need to return the first row.
Returning sequence vectors as rows of a sorted
matrix limits the number of possible sequences to
the matrix size. This can be extended if we define
a similarity threshold T and return a random index
Procedure GET-ONSET-LIST
for n in GridSize do
R = Random number 0 < Variance
I = Index from Row R of Similarity
Matrix
S = New SoundConnection
S->SoundUnit = SoundUnit(I)
Add S to LinkedList
end for
return LinkedList
End Procedure
entre 0 and j − T for each step i in the new
secuencia.
When the user presses the New Pattern button
(Figure 2c), a new linked list of objects, called sound
connections, is formed. This represents a traversal
through connected sound objects in the timbre space.
The length of the linked list is determined by the grid
size specified by the user, so if the user specifies, para
ejemplo, a grid size of 1/16, a one-measure sequence
of 16th notes will be generated. The algorithm in
Cifra 3 details the exact procedure whereby we
generate the list. The variance parameter affects the
threshold of similarity by which onsets are chosen.
Con 0 variance, the most similar sequence is always
returned. This variance parameter is adjustable from
the Accuracy/Variety slider in the lower-left corner
of the instrument (Figure 2c).
In the main timbre space interface (Figure 2d),
a visual graph is generated in the timbre space by
traversing the linked list and drawing line edges
connecting each sound object pointed to by the
sound connection in the linked list. En este caso, a
loop of 16 onsets has been generated, with the onset
numbers indicated beside the associated sound
object for each onset in the sequence. The user
is free to manipulate these sound connections to
mutate these patterns by touching or clicking on
the sound connection and dragging to another sound
object. Multiple sound connections assigned to an
individual sound object can be selected as a group
by slowly double-tapping and then dragging.
On the audio side, every time there is a new beat,
the linked list is traversed. If a sound connection’s
onset number matches the current beat, the corre-
sponding sound unit is played back. One addition
´O Nuan ´ain et al.
29
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
that occurred after some user experiments with the
prototype is the linear waveform representation of
the newly generated sequence (Figure 2e). Users felt
the combination of the 2-D interface with the tradi-
tional waveform representation made the sequences
easier to navigate and they also welcomed being able
to manipulate the internal arrangement of sequence
itself once generated.
Evaluation
In the course of our literature review of the state of
the art, we were particularly interested in examining
the procedures and frameworks used in performing
evaluations of the implemented systems. Our most
immediate observation was that evaluation is an
understudied aspect of research into concatenative
sistemas. With creative and generative systems,
this is often the case; many such systems are
designed solely with the author as composer in
mente.
Some authors provide examples of use cases
(Cardle, Arroyos, and Robinson 2003). Autores, semejante
as Sturm, have made multimedia examples available
on the Web (see Zils and Pachet 2001; Xiang 2002;
Sturm 2004). Frequently, researchers have made
allusions to some concept of the “user,” but only
one paper has presented details of a user experiment
(Aucouturier and Pachet 2005). One researcher,
Graham Coleman, also highlighted this lack of
evaluation strategies in concatenative synthesis
in his doctoral dissertation (Coleman 2015). Para
the evaluation of his own system, he undertook a
listening experiment with human participants in
tandem with a thorough analysis of algorithmic
performance and complexity.
We conducted extensive evaluation of our own
sistema, both quantitatively and qualitatively. En
the quantitative portion, we set out to investigate
two key aspects. Primero, if we consider the system as
a retrieval task that aims to return similar items,
how accurate and predictable is the algorithm and
its associated distance metric? Segundo, how does
this objective retrieval accuracy correspond to the
perceptual response of the human listener to the
retrieved items?
The qualitative evaluation consisted of inter-
active, informal interviews with intended users—
mostly active music producers but also music
researchers and students—as they used the soft-
mercancía. We gathered their responses and impressions
and grouped them according to thematic analysis
técnicas. As alluded to in the introduction, ambos
the quantitative evaluation and the qualitative eval-
uation have been previously reported in separate
publicaciones, but we include summaries of each here
for reference.
System Evaluation
We describe here the qualitative portion of the
evaluación, first by introducing the experimental
setup, then presenting and comparing the results of
the algorithm’s retrieval accuracy with the listener
survey.
Experimental Setup
Because the goal of the system is the generation
of rhythmic loops, we decided to formulate an
experiment using breakbeats (short drum solos taken
from commercial funk and soul stereo recordings).
Ten breakbeats were chosen in the range 75–
142 bpm, and we truncated each of them to a
single bar in length. Repeating ten times for each
loop, we selected a single loop as the target seed
and resynthesized it using the other nine loops
(similar to holdout validation in machine learning)
at four different distances from target to create 40
variations.
Each of the loops was manually labeled with
the constituent drum sounds as we hear them.
The labeling used was “K” for kick drum, “S” for
snare, “HH” for hi-hat, “C” for cymbal, and “X”
when the content was not clear (such as artifacts
from the onset-detection process or some spillage
from other sources in the recording). Cifra 4 muestra
the distribution labels in the entire data set and
the distribution according to step sequence. Nosotros
can notice immediately the heavy predominance
of hi-hat sounds, which is typical in kit-based
drumming patterns. Además, the natural trends
30
Computer Music Journal
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
Cifra 4. Distribution of
sound labels in the source
cuerpo (a). Distribution of
sound labels by step
number in the 16-step
secuencia (b).
Cifra 5. Scatter plot and
linear regression of
accuracy versus distance
for all sound labels (a) y
for the same sequence
with kick drum and snare
isolated (b).
yo
D
oh
w
norte
oh
a
d
mi
d
F
r
oh
metro
h
t
t
pag
:
/
/
d
i
r
mi
C
t
.
metro
i
t
.
mi
d
tu
/
C
oh
metro
j
/
yo
a
r
t
i
C
mi
–
pag
d
F
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
C
oh
metro
_
a
_
0
0
4
1
2
pag
d
.
j
F
b
y
gramo
tu
mi
s
t
t
oh
norte
0
7
S
mi
pag
mi
metro
b
mi
r
2
0
2
3
of kit drumming are evident, a saber, kick-drum
placement on the first beat and offbeat peaks for the
snares.
Retrieval Evaluation
We compared each of the labels in each 16-step
position of the quantized target loop with the labels
in each step of the newly generated sequences. El
accuracy A of the algorithm is then given by the
number of correctly retrieved labels divided by the
total number of labels in the target loop, inspired
by a similar approach adopted by Thompson, dixon,
and Mauch (2014).
Based on Pearson’s correlation of the retrieval
ratings and the distances of the generated patterns,
we were able to confirm the tendency of smaller
distances to produce more similar patterns in
terms of the labeling accuracy. A moderate
negative correlation of r = −0.516 (significance
level p < 0.001) is visible by the regression line
in Figure 5a. If we isolate the kick and snare
(often considered the more salient events in
drum performances, see Gouyon, Pachet, and
Delerue 2000) the negative correlation value
decreases sharply to r = −0.826, as shown in
Figure 5b.
´O Nuan ´ain et al.
31
Listener Evaluation
Observing that the algorithm tends to reproduce
labels in a predictable fashion, we sought to establish
whether this conforms in reality to what a human
listener perceives. An online listener survey was
conducted using the same generated loops and
targets from the retrieval evaluation. Twenty-one
participants completed the survey, drawn mostly
from music researchers and students from the
institutions of Universitat Pompeu Fabra and
the Escola Superior de M ´usica de Catalunya in
Barcelona, as well as friends with an interest in
music. Twenty out of those indicated that they
played an instrument, with nine specifying an
instrument from the percussion family.
The participants were requested to audition
a target loop and each of the generated loops in
succession. They were then asked to rate, on a
Likert scale of 1 to 5, the similarity of the generated
loop to the target in terms of their timbre (i.e., do
the kick drums, snares, and hi-hats sound alike?) as
well as the rhythmic structure of the pattern (i.e.,
is the arrangement and placement of the sounds
similar?). We also asked them to rate their aesthetic
preference for the generated patterns, to determine
any possible correlation with similarity.
Survey results were collated and analyzed using
Spearman’s rank correlation, comparing the mode
of the participants’ responses with the distance
value of each loop. A moderate-to-strong negative
correlation pattern emerged for all of the variables
under consideration, namely, r = −0.66 for pattern
similarity, r = −0.59 for timbral similarity, and
r = −0.63 for their personal preference according to
similarity (with significance levels of p < 0.01 in all
instances). It should be evident that the listeners’
judgments reflect what the results unearthed in the
retrieval evaluation.
User Evaluation
ing of evaluative scrutiny is the users’ experience of
working with the software: gauging their responses
to the interface, its modes of interactions, and its rel-
evance and suitability for their own compositional
styles and processes.
To this effect, a qualitative evaluation phase was
arranged to gather rich descriptive impressions from
related groups of users in Barcelona and Berlin during
February 2016. In Barcelona, as with user profiles
of the listener survey, most of the participants were
researchers or students in the broad area of Sound
and Music Computing. In Berlin we were able to
gain access to artists involved in the Red Bull Music
Academy as well as with employees of the music
software company Native Instruments.
In broad terms, the overall sense of people’s
impressions was positive. Many participants were
initially attracted to the visual nature of the soft-
ware and were curious to discover its function
and purpose. After some familiarization with its
operation, people also remarked positively on its
sonic output and ability to replicate the target
loop:
“It’s an excellent tool for making small changes
in real time. The interface for me is excellent.
This two-dimensional arrangement of the
different sounds and its situation by familiarity,
it’s also really good for making these changes.”
“I’m really interested in more-visual, more-
graphical interfaces. Also, the fact that you can
come up with new patterns just by the push of
a button is always great.”
“It’s inspiring because this mix makes some-
thing interesting still, but also I have the feeling
I can steal it.”
“The unbelievable thing is that it can create
something that is so accurate. I wouldn’t believe
that it’s capable of doing such a thing.”
The quantitative evaluation demonstrated the pre-
dictive performance of the algorithm based on
retrieval accuracy and the corresponding listeners’
judgments of similarity and likeness. Equally deserv-
Some of the negative criticism came from the
prototypical nature of the instrument, and some
users were not comfortable with its perceived
indeterminacy:
32
Computer Music Journal
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
c
o
m
j
/
l
a
r
t
i
c
e
-
p
d
f
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
c
o
m
_
a
_
0
0
4
1
2
p
d
.
j
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
“It was too intense and also [had] a prototype
feeling. So I was like, ‘Well, it’s cool and very
interesting but not usable yet.’ ”
“Right now it’s still hard to find your way
around, but that’s something you can refine
pretty easily.”
Usage Scenarios
Participants were asked to consider how they
would envisage themselves using the software.
Most of them concurred that its strength would
be in supporting musicians in their production and
compositional workflows. Some users were curious
about using it in live contexts, such as continuous
analysis of instrumental performance or beat-boxing
assistance:
“This is great! Ah, but wait . . . Does it mean
I could, like, beat box really badly some idea
that I have . . . and then bring my samples, my
favorite kits and then it will just work?”
Continuous recording and analysis is within
the realm of possibility, but can potentially be an
operation that is prohibitively computationally
expensive, depending on the granularity of the beat
grid and the size of the corpus. Further benchmarking
and tests are required to establish the upper bounds.
Another interesting observation was that many
users did not want to start with a target, preferring
to use the instrument as a new, systematic method
of exploring their existing sounds:
“I’ve got this fully on wet straight away, which
tells you the direction I’d be going with it.”
“You just want to drag in a hundred different
songs and you just want to explore without
having this connection to the original group.
Just want to explore and create sound with it.”
adapt the new visual paradigm, they still felt the need
for a linear waveform to aid their comprehension.
Because of this feedback, the waveform view was
implemented early on in our development, as is
evident in its inclusion in Figure 2.
“It’s a bit hard to figure out which sixteenth
you are looking for, because you are so used to
seeing this as a step grid.”
‘You have a waveform or something. . . Then I
know, okay, this is the position that I’m at.”
“Is there also a waveform place to put the
visualization? People are so used to having that
kind of thing.”
Shaping Sounds
A recurring issue, which cropped up mainly with
producers and DJs, was the desire to shape, process,
and refine the sounds once a desirable sequence was
generated by the system. This way of composing
seems emblematic of electronic music producers
across the board; they start with a small loop or idea
then vary and develop it exploiting the many effects
processing and editing features provided by their
tools. Most crucially, they desired the option to be
able to control the envelopes of the individual units
via drawable attack and decay parameters, which is
currently being implemented.
“. . . an attack and decay just to sort of tighten
it up a little bit . . . get rid of some of the rough
edges of the onsets and offsets.”
“It would be great if you could increase the
decay of the snare, for example. Which, if it’s
prototype, you can’t expect to have all those
functions there immediately, but in an end
product, I think it would be a necessity.”
Traditional Forms of Navigation
Our original intention was for users to solely be able
to arrange their patterns through the 2-D timbre
space. Through the course of our discussions with
users we learned that, although they were eager to
Parameterization and Visualization
The most overarching source of negative criti-
cism from all users was in how we presented
the parameters of the system. Users are freely
able to manipulate the individual weightings of
the features, affecting their relative influence in
´O Nuan ´ain et al.
33
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
c
o
m
j
/
l
a
r
t
i
c
e
-
p
d
f
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
c
o
m
_
a
_
0
0
4
1
2
p
d
.
j
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
the similarity computation, but also in the PCA
dimensional-reduction stage. In an effort to make
this more “user friendly,” we relabeled the fea-
ture names with more generally comprehensible
terms like “timbre,” “brightness,” “harmonicity,”
and “loudness.” Despite this, participants reported
being confused and overwhelmed by this level of
control, stating that they were a “a bit lost already,”
that there are “four parameters, and you don’t know
which thing is what,” and that they “would prefer
not to have too many controls.”
Most users were quite content with the overall
sonic output from the system without delving into
the manipulation of feature parameters. For the
visualization, however, there are certain configura-
tions of the features that produce the best separation
and clustering of the units (although MFCCs alone
appear to be the most robust in our experience).
One option we are actively investigating would
be to remove these parameter sliders and replace
them with an optional “advanced” mode, giving
users the ability to select specific parameters for
the axes (as in CataRT) in addition to “automatic”
arrangement configurations made possible by using
dimensionality-reduction techniques. These con-
figurations could be derived by analyzing different
sound sets to find weighting combinations that
give the best visual separation, depending on the
corpus provided. Finally, we are currently using PCA
for dimensionality reduction. There are also other
approaches, including multidimensional scaling
(Donaldson, Knopke, and Raphael 2007) and the
recent t-distributed stochastic neighbor embedding
algorithm (Frisson 2015; Turquois et al. 2016), which
have been used in musically related tasks and that
we are implementing and evaluating as alternatives.
Discussion
Evaluating systems for music creation and manip-
ulation is a difficult, ill-defined, and insufficiently
reported task. As we have stressed in the course of
this article, this is also the case with systems for
concatenative synthesis. After conducting our own
evaluation, we considered what key points could be
made to help inform future evaluations by interested
researchers in the community. Our observations led
us to indicate three distinct layers that should be
addressed for a significant, full-fledged appraisal.
The most high-level and general “system” layer
calls for user evaluations that go beyond “quality of
experience” and “satisfaction” surveys. Such evalu-
ations should strive to address creative productivity
and workflow efficiency aspects particular to the
needs of computer-music practitioners.
At the mid-level “algorithmic” layer, we examine
the mechanics of developing solutions strategies for
concatenative synthesis. We have identified three
main trends in algorithmic techniques used for
tackling tasks in concatenative synthesis, namely,
similarity-matrix and clustering approaches (like
ours), Markov models, and constraint-satisfaction
problems. Each of these techniques exhibits its own
strengths and weaknesses in terms of accuracy,
flexibility, efficiency, and complexity. Comparing
these algorithms within a single system and, indeed,
across multiple systems, using a well-defined
data set, a clear set of goals, and specific success
criteria would represent a valuable asset in the
evaluation methodology of concatenative synthesis.
Additionally, we should pay attention to the distance
and similarity metrics used, as there are other
possibilities that are explored and compared in other
retrieval problems (e.g., Charulatha, Rodrigues, and
Chitralekha 2013).
At the lowest level, the focus is on the broader
implications related to MIR of choosing appropriate
features for the task at hand. In the course of our
evaluation, we chose the features indicated in the
implementation and did not manipulate them in
the experiment. There are, of course, many other
features relevant to the problem that can be studied
and estimated in a systematic way, as is par for
the course in classification experiments in MIR.
Furthermore, tuning the weights was not explored
and is an important consideration that depends
greatly on different corpora and output-sequence
requirements.
In addition to this three-tiered evaluation method-
ology, an ideal component would be the availability
of a baseline or comparison system that ensures new
prototypes improve over some clearly identifiable
aspect. Self-referential evaluations run the risk of
34
Computer Music Journal
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
c
o
m
j
/
l
a
r
t
i
c
e
-
p
d
f
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
c
o
m
_
a
_
0
0
4
1
2
p
d
.
j
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
confirming experimenter bias without establishing
comprehensive criticism.
References
Conclusion
In this article, we explored concatenative synthesis
as a compositional tool for generating rhythmic
patterns for electronic music, with a strong empha-
sis on its role in EDM musical styles. One of our
first contributions was to present a thorough and
up-to-date review of the state of the art, beginning
with its fundamental algorithmic underpinnings
and proceeding to modern systems that exploit new
and experimental visual and interactive modalities.
Although there are a number of commercial applica-
tions that encapsulate techniques of concatenative
synthesis for the user, the vast majority of systems
are frequently custom-built for the designer or are
highly prototypical in nature. Consequently, there
is a marked lack of evaluation strategies or reports
of user experiences in the accompanying literature.
Based on these investigations, we set out to
design a system that applied and extended many of
the pervasive techniques in concatenative synthesis
with a clear idea of its application and its target
user. We built an instrument that was easily
integrated with modern digital audio workstations
and presented an interface that intended to be
attractive and easy to familiarize oneself with.
How to evaluate the system, not only in terms of
its objective performance but also in its subjective
aural and experiential implications for our users, was
our final substantial contribution to this area. The
results of our evaluations showed that our system
performed as expected, and users were positive
about its potential for assisting in their creative
tasks, while also proposing interesting avenues for
future work and contributions.
Resources
A demonstration version of the software is available
online at http://github.com/carthach/rhythmCAT.
A video example can be viewed at http://youtu.be
/hByhgF fzto.
Aucouturier, J.-J., and F. Pachet. 2005. “Ringomatic:
A Real-Time Interactive Drummer Using Constraint-
Satisfaction and Drum Sound Descriptors.” Proceedings
of the International Conference on Music Information
Retrieval, pp. 412–419.
Bart ´ok, B. 1993. “Hungarian Folk Music.” In B. Suchoff,
ed. B ´ela Bart ´ok Essays. Lincoln: University of Nebraska
Press, pp. 3–4.
Battier, M. 2007. “What the GRM Brought to Music: From
Musique Concr `ete to Acousmatic Music.” Organized
Sound 12(3):189–202.
Bello, J., and L. Daudet. 2005. “A Tutorial on Onset
Detection in Music Signals.” IEEE Transactions on
Audio, Speech, and Language Processing 13(5):1035–
1047.
Bernardes, G., C. Guedes, and B. Pennycook. 2013.
“EarGram?: An Application for Interactive Exploration
of Concatenative Sound Synthesis in Pure Data.” In
M. Aramaki et al., eds. From Sounds to Music and
Emotions. Berlin: Springer, pp. 110–129.
B ¨ock, S., and G. Widmer. 2013. “Maximum Fil-
ter Vibrato Suppression for Onset Detection.”
In Proceedings of the International Conference
on Digital Audio Effects. Available online at
dafx13.nuim.ie/papers/09.dafx2013 submission 12.pdf.
Accessed January 2017.
Bogdanov, D., et al. 2013. “ESSENTIA: An Audio Analysis
Library for Music Information Retrieval.” In Pro-
ceedings of the International Conference on Music
Information Retrieval, pp. 493–498.
Bradski, G. 2000. “The OpenCV Library.” Dr. Dobb’s
Journal 25(11):120–125.
Brent, W. 2010. “A Timbre Analysis and Classifica-
tion Toolkit for Pure Data.” In Proceedings of the
International Computer Music Conference, pp. 224–
229.
Cardle, M., S. Brooks, and P. Robinson. 2003. “Audio
and User Directed Sound Synthesis.” Proceedings of
the International Computer Music Conference, pp.
243–246.
Charulatha, B., P. Rodrigues, and T. Chitralekha. 2013.
“A Comparative Study of Different Distance Metrics
That Can Be Used in Fuzzy Clustering Algorithms.”
International Journal of Emerging Trends and Tech-
nology in Computer Science. Available online at
www.ijettcs.org/NCASG-2013/NCASG 38.pdf. Ac-
cessed January 2017.
Coleman, G. 2015. “Descriptor Control of Sound Transfor-
mations and Mosaicing Synthesis.” PhD dissertation,
´O Nuan ´ain et al.
35
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
c
o
m
j
/
l
a
r
t
i
c
e
-
p
d
f
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
c
o
m
_
a
_
0
0
4
1
2
p
d
.
j
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Universitat Pompeu Fabra, Department of Information
and Communication Technologies, Barcelona.
Dixon, S. 2006. “Onset Detection Revisited.” In Proceed-
ings of the International Conference on Digital Audio
Effects, pp. 133–137.
Hunt, A. J., and A. W. Black. 1996. “Unit Selection in
a Concatenative Speech Synthesis System Using a
Large Speech Database.” In Proceedings of the IEEE
International Conference on Acoustics, Speech, and
Signal Processing, pp. 373–376.
Donaldson, J., I. Knopke, and C. Raphael. 2007. “Chroma
Jehan, T. 2005. “Creating Music by Listening.” PhD
Palette?: Chromatic Maps of Sound as Granular
Synthesis Interface.” In Proceedings of the Conference
on New Interfaces for Musical Expression, pp. 213–219.
Frisson, C. 2015. “Designing Interaction for Browsing
Media Collections (by Similarity).” PhD dissertation,
Universit ´e de Mons, Faculty of Engineering.
Frisson, C., C. Picard, and D. Tardieu. 2010. “Audiog-
arden?: Towards a Usable Tool for Composite Audio
Creation.” QPSR of the Numediart Research Program
3(2):33–36.
Gouyon, F., F. Pachet, and O. Delerue. 2000. ”On the
Use of Zero-Crossing Rate for an Application of
Classification of Percussive Sounds.” In Proceedings of
the International Conference on Digital Audio Effects,
pp. 3–8.
Greenberg, C. 1971. “Collage.” In Art and Culture:
Critical Essays. Boston, Massachusetts: Beacon Press,
pp. 70–83.
Hackbarth, B., N. Schnell, and D. Schwarz. 2011. “Au-
dioGuide?: A Framework for Creative Exploration
of Concatenative Sound Synthesis.” IRCAM Re-
search Report. Available online at articles.ircam.fr
/textes/Hackbarth10a/index.pdf. Accessed January
2017.
Herrera, P., A. Dehamel, and F. Gouyon. 2003. “Au-
tomatic Labeling of Unpitched Percussion Sounds.”
In Proceedings of the 114th Audio Engineering So-
ciety Convention. Available online at www.aes.org
/e-lib/browse.cfm?elib=12599 (subscription required).
Accessed January 2017.
Holmes, T. 2008. Electronic and Experimental Music.
Abingdon-on-Thames, UK: Routledge.
Hoskinson, R., and D. Pai. 2001. “Manipulation and
Resynthesis with Natural Grains.” In Proceedings
of the International Computer Music Conference,
pp. 338–341.
Hoskinson, R., and D. K. Pai. 2007. “Synthetic Sound-
scapes with Natural Grains.” Presence: Teleoperators
and Virtual Environments 16(1):84–99.
Humphrey, E. J., D. Turnbull, and T. Collins. 2013. “A
Brief Review of Creative MIR.” In Proceedings of
the International Conference on Music Information
Retrieval. Available online at ismir2013.ismir.net
/wp-content/uploads/2014/02/lbd1.pdf. Accessed Jan-
uary 2017.
dissertation, Massachusetts Institute of Technology,
Media Arts and Sciences.
Knees, P., et al. 2016. “The GiantSteps Project: A Second-
Year Intermediate Report.” In Proceedings of the
International Computer Music Conference, pp. 363–
368.
Lochhead, J. I., and J. H. Auner. 2002. Postmodern Music/-
Postmodern Thought: Studies in Contemporary Music
and Culture. Abingdon-on-Thames, UK: Routledge.
Logan, B. 2000. “Mel Frequency Cepstral Coefficients for
Music Modeling.” In Proceedings of the International
Symposium on Music Information Retrieval. Available
online at ismir2000.ismir.net/papers/logan paper.pdf.
Accessed January 2017.
Loughran, R., et al. 2004. “The Use of Mel-Frequency
Cepstral Coefficients in Musical Instrument Identifi-
cation.” In Proceedings of the International Computer
Music Conference, pp. 42–43.
Mangani, M., R. Baldizzone, and G. Nobile. 2006. “Quo-
tation in Jazz Improvisation: A Database and Some
Examples.” Paper presented at the International Con-
ference on Music Perception and Cognition, August
22–26, Bologna, Italy.
Masri, P., and A. Bateman. 1996. “Improved Modeling
of Attack Transients in Music Analysis-Resynthesis.”
In Proceedings of the International Computer Music
Conference, pp. 100–103.
McLeod, K. 2009. “Crashing the Spectacle: A Forgotten
History of Digital Sampling, Infringement, Copyright
Liberation and the End of Recorded Music.” Culture
Machine 10:114–130.
Metzer, D. 2003. Quotation and Cultural Meaning in
Twentieth-Century Music. Cambridge: Cambridge
University Press.
Miller, P. D. 2008. Sound Unbound: Sampling Digital
Music and Culture. Cambridge, Massachusetts: MIT
Press.
O’Connell, J. 2011. “Musical Mosaicing with High Level
Descriptors.” Master’s thesis, Universitat Pompeu
Fabra, Sound and Music Computing, Barcelona.
´O Nuan ´ain, C., P. Herrera, and S. Jord `a. 2016. “An
Evaluation Framework and Case Study for Rhythmic
Concatenative Synthesis.” In Proceedings of the
International Society for Music Information Retrieval
Conference, pp. 67–72.
36
Computer Music Journal
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
c
o
m
j
/
l
a
r
t
i
c
e
-
p
d
f
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
c
o
m
_
a
_
0
0
4
1
2
p
d
.
j
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
´O Nuan ´ain, C., S. Jord `a, and P. Herrera. 2016a. “An In-
teractive Software Instrument for Real-Time Rhythmic
Concatenative Synthesis.” In Proceedings of the In-
ternational Conference on New Interfaces for Musical
Expression, pp. 383–387.
´O Nuan ´ain, C., S. Jord `a, and P. Herrera. 2016b. “Towards
User-Tailored Creative Applications of Concatena-
tive Synthesis in Electronic Dance Music.” In Pro-
ceedings of the International Workshop on Musical
Metacreation. Available online at musicalmetacre-
ation.org/buddydrive/file/nuanain towards. Accessed
January 2017.
Oswald, J. 1985. “Plunderphonics, or Audio Piracy as a
Compositional Prerogative.” Paper presented at the
Wired Society Electro-Acoustic Conference, Toronto,
Canada. Reprinted in Musicworks, Winter 1986, 34:5–8.
Peeters, G. 2004. “A Large Set of Audio Features for
Sound Description (Similarity and Classification)
in the CUIDADO Project.” IRCAM Project Report.
Available online at recherche.ircam.fr/anasyn/peeters
/ARTICLES/Peeters 2003 cuidadoaudiofeatures.pdf.
Accessed January 2017.
Rabiner, L. R. 1989. “A Tutorial on Hidden Markov Models
and Selected Applications in Speech Recognition.”
Proceedings of the IEEE 77(2):257–286.
Roads, C. 2004. Microsound. Cambridge, Massachusetts:
MIT Press.
Rodgers, T. 2003. “On the Process and Aesthetics of
Sampling in Electronic Music Production.” Organized
Sound 8(3):313–320.
Roy, P., F. Pachet, and S. Krakowski. 2007. “Analytical
Features for the Classification of Percussive Sounds:
The Case of the Pandeiro.” In Proceedings of the
International Conference on Digital Audio Effects,
pp. 213–220.
Russell, S., and P. Norvig. 2009. Artificial Intelligence:
A Modern Approach. Upper Saddle River, New Jersey:
Prentice Hall.
Schwarz, D., et al. 2006. “Real-Time Corpus-Based
Concatenative Synthesis with CataRT.” In Proceedings
of the International Conference on Digital Audio
Effects, pp. 18–21.
Sewell, A. 2014. “Paul’s Boutique and Fear of a Black
Planet: Digital Sampling and Musical Style in Hip
Hop.” Journal of the Society for American Music
8(1):28–48.
Sturm, B. L. 2004. “Matconcat: An Application for Explor-
ing Concatenative Sound Synthesis Using Matlab.” In
Proceedings of the International Conference on Digital
Audio Effects, pp. 323–326.
Sturm, B. L. 2006. “Adaptive Concatenative Sound Synthe-
sis and Its Application to Micromontage Composition.”
Computer Music Journal 30(4):46–66.
Tamagawa, K. 1988. “Echoes from the East: The Javanese
Gamelan and Its Influence on the Music of Claude
Debussy.” DMA dissertation, University of Texas at
Austin.
Thompson, L., S. Dixon, and M. Mauch. 2014. “Drum
Transcription via Classification of Bar-Level Rhythmic
Patterns.” In Proceedings of the International Society
for Music Information Retrieval Conference, pp. 187–
192.
Tindale, A., et al. 2004. “Retrieval of Percussion Gestures
Using Timbre Classification Techniques.” Proceedings
of the International Conference on Music Information
Retrieval, pp. 541–545.
Turquois, C., et al. 2016. “Exploring the Benefits of
2D Visualizations for Drum Samples Retrieval.” In
Proceedings of the ACM SIGIR Conference on Human
Information Interaction and Retrieval, pp. 329–
332.
Tzanetakis, G., G. Essl, and P. Cook. 2001. “Automatic
Musical Genre Classification of Audio Signals.” In
Proceedings of the International Symposium on Music
Information Retrieval, pp. 293–302.
Xenakis, I. 1971. Formalized Music. Bloomington: Indiana
Schwarz, D. 2003. “The Caterpillar System for Data-
University Press.
Driven Concatenative Sound Synthesis.” In Proceedings
of the International Conference on Digital Audio
Effects, pp. 135–140.
Schwarz, D. 2005. “Current Research in Concatenative
Sound Synthesis.” In Proceedings of the International
Computer Music Conference, pp. 9–12.
Schwarz, D. 2017. “Corpus-Based Sound Synthesis
Survey.” Available online at imtr.ircam.fr/imtr/Corpus
-Based Sound Synthesis Survey. Accessed February
2017.
Xiang, P. 2002. “A New Scheme for Real-Time Loop
Music Production Based on Granular Similarity and
Probability Control.” In Proceedings of the Interna-
tional Conference on Digital Audio Effects, pp. 89–
92.
Zils, A., and F. Pachet. 2001. “Musical Mosaicing.”
In Proceedings of the International Conference
on Digital Audio Effects, 1–6. Available online at
www.csis.ul.ie/dafx01/proceedings/papers/zils.pdf.
Accessed January 2017.
´O Nuan ´ain et al.
37
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
c
o
m
j
/
l
a
r
t
i
c
e
-
p
d
f
/
/
/
/
4
1
2
2
1
1
8
5
6
3
9
3
/
c
o
m
_
a
_
0
0
4
1
2
p
d
.
j
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
