Perspectives on Real-time Computation
of Movement Coarticulation
Ircam - Centre Pompidou
Montreal, QC, Canada
School of Interactive Arts
Simon Fraser University
We discuss the notion of movement coarticulation, which has
been studied in several ﬁelds such as motor control, music
performance and animation. In gesture recognition, move-
ment coarticulation is generally viewed as a transition be-
tween “gestures” that can be problematic. We propose here
to account for movement coarticulation as an informative ele-
ment of skilled practice and propose to explore computational
modeling of coarticulation. We show that established proba-
bilistic models need to be extended to accurately take into
account movement coarticulation, and we propose research
questions towards such a goal.
Coarticulation; Movement; Gesture; Recognition; Motor
ACM Classiﬁcation Keywords
H.5. Information Interfaces and Presentation (e.g. HCI):
Multimedia Information Systems; G.3Probability And Statis-
tics: Time series analysis; J.5 Arts and Humanities: Perform-
ing arts (e.g., dance, music)
Coarticulation is a well known phenomena in speech produc-
tion occurring when a sound segment is inﬂuenced by its con-
text, such as the preceding and following sound segments in
a word or sentence1. This problem has been widely studied
and modeled for both speech recognition and generation .
While coarticulation has also been described in movement se-
quences, it remains largely overlooked. In Human-Computer
Interaction (HCI), and particularly in gesture-based interac-
tion, the phenomenon of coarticulation is often considered as
1We note that the word coarticulation is also used to describe the oc-
currence of two different modalities, for example voice and gesture.
In this paper, we used the term coarticulation as it is generally used
in speech production.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for proﬁt or commercial advantage and that copies bear this notice and the full citation
on the ﬁrst page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise,
or republish, to post on servers or to redistribute to lists, requires prior speciﬁc
permission and/or a fee. Request permissions from Permissions@acm.org.
MOCO’16, July 05 - 07, 2016, Thessaloniki, GA, Greece
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
a “problem”, since it perturbs the performance of gestural vo-
cabularies and can reduce recognition rate .
Coarticulation relies on the existence and formalization of
constitutive segments. For instance speech is often examined
as a ﬁnite number of phonological sound segments that are
ordered, and which ordering is linguistic-dependent. Coar-
ticulation can be observed and measured because the alter-
ation of phonemes remains consistent over a large vocabu-
lary. Considering movement, such segments become highly
variable across individuals and context-dependent, making
their formalization inherently more complex. Motor theo-
rists proposed the notion of movement primitives  as basic
units (typically, patterns of movement kinematics) that can
be sequenced to execute a complex movement. Within this
framework, movement coarticulation has been linked to mo-
tor skills that involve the selection of movement primitives,
their ordering and their accurate execution .
In this context, an important challenge is the computational
modeling of coarticulating movements with the aim of lever-
aging on user’s motor skills rather than discarding them for
the sake of accuracy in gesture recognition systems. This pa-
per aims at discussing problems and prospects with compu-
tational modeling of coarticulation for the ﬁeld of movement
and computing. In particular, we propose to include aspects
from motor control theories. We ﬁrst recall some important
references for movement coarticulation that span over differ-
ent disciplines. Second we present computational approaches
in interactive systems. We then illustrate typical coarticula-
tion phenomena occurring with simple gestural inputs and
the analysis of these phenomena through the lens of existing
computational models. Finally we shortly discussed the re-
sults and propose a perspective for computational movement
models involving coarticulation.
COARTICULATION ACROSS DIFFERENT FIELDS
Coarticulation has been studied and formalized in speech per-
ception  and recognition [25, 4], communication, anima-
tion, embodied conversational agents  as well as sign lan-
guage (both in pure synthesis and motion retrieval and data-
driven sign-language synthesis) [11, 26, 17].
In motor control, coarticulation has been studied considering
simple tasks and movements [15, 27, 29, 23]. Typically, coar-
ticulation occurs in sensori-motor learning when the move-
ment primitives (understood as basic units such as patterns of
movement kinematics ) are fused in a larger phrase, which
does not appear as a series of separate events , but from an
intermediate-level command that encompasses several events
as one event, called “chunk”. Chunks’ boundaries are char-
acterised by higher motor variability and probability of er-
rors. This behavioral characterisation is supported by recent
work in neuroscience proposing a hierarchical motor repre-
sentation underlying expert performance (see for example the
recent review of ). As a corollary, coarticulation is gen-
erally considered as the results of an anticipative behaviour:
the execution of the unit is planned ahead and the movement
appears to start before the end of the previous unit .
The concept of coarticulation has also found an echo in the
study of musical gestures, and in particular instrumental ges-
tures [13, 20, 22, 2, 12, 20]. In particular, Godøy  pro-
poses an informative review of coarticulation in music perfor-
mance. In music performance, the notion of small movement
units can be linked to existing musical events such as musical
notes or sounds. Such link between score events and seg-
mented instrumentalists’ gestures have also been examined
through the use of computational model . Motor theories
and cognition should be considered here since instrumental
gestures imply learning skilled movements and anticipation.
Yet, important works remains to be conducted on this area.
Finally, in dance, motor skill learning is acknowledged as a
fundamental aspect of the practice. Thus dance constitutes
also a promising ﬁeld for investigating coarticulation. Nev-
ertheless, to our knowledge the notion of coarticulation in
dance has been less studied from a computational perspec-
tive. Here the challenges are two-fold: to deﬁne what consti-
tutes a movement segment that can then be used in comput-
ing systems, and to design computing systems able to under-
stand higher-level representations of complex dance move-
ments coherent with embodied cognitive mechanisms .
REAL-TIME COMPUTATION OF COARTICULATION
In this work, we consider gesture-based interactive systems in
which movement analysis must be achieved in real-time. In
such cases, the challenge is to identify and characterize seg-
ments from a continuous stream of motion data while simul-
taneously accounting for their context-dependent variations.
In this section, we introduce the type of models that we are
considering for real-time computation of coarticulation.
In the literature, spotting and classifying gestures in a contin-
uous stream of movement data is usually called continuous
gesture recognition. State-space temporal models are typi-
cally used for continuous gesture recognition because they
can take into account both variability in execution and tempo-
ral dependencies in the signal . For example, Conditional
Random Fields (CRFs) have been shown successful for such
a task [32, 18]. Moreover, considering coarticulation, CRF
is able to take into account contextual information such as
the preceding and following of a given segment. However,
as described in Morency et al. , standard implementa-
tion of CRF can only be used for ofﬂine analysis on bounded
continuous stream preventing for its use in interactive sys-
tems. Nevertheless, Song et al  extended the models by
proposing an online spotting within a sliding window, but not
addressing explicitly articulation effects.
Considering gesture-based interactive systems, we believe
that the challenge is actually to go beyond continuous gesture
recognition (i.e.spotting) in characterizing the gestures exe-
cution and in particular their coarticulation . Importantly,
physical movements are dynamic phenomena, which encode
features directly linked to expressivity. It is particularly true
for coarticulatory movements that are prone to unconscious
cognition-induced changes in dynamics (e.g. chunking) as
well as voluntary (conscious) continuous variations.
In our previous work, we have proposed Bayesian state-space
models able to infer in real-time the gesture performed and
characteristics of its execution. We have proposed two main
approaches for this problem: a template-based continuous
state-space model , and a variant of hidden Markov models
[10, 9]. The former is able to track modulations of recorded
templates, the latter is able to learn statistically-relevant ges-
ture variances. As both approaches proposed two comple-
mentary views of potential variability in movement execu-
tion, we inspect in the next section how these models can
inform on the coarticulatory content in gesture sequences.
A TOY EXAMPLE
This section aims to illustrate a typical case of gesture coar-
ticulation and the associated challenges for real-time analysis
and continuous gesture recognition.
We recorded a set of executions of two gestures drawn on
a trackpad, using Cycling’74 Max2with the external ﬁnger-
pinger3to measure the trajectories. We also used the library
MuBu4for Max  in order to record and save the captured
We chose to use two gestures, typically used in gesture-based
interactive systems (see for example ). The ﬁrst gesture
is a “V” (Figure 1, a) the second is a “O” (Figure 1, b). These
two gestures are then sequenced in two different ways: Ges-
ture 1– Stop – Gesture 2and Gesture 1– Gesture 2(no pause
between both gesture executions). The 4gestures are depicted
in Figure 1. Each one of these gestures is repeated 10 times.
The two different ways to perform transitions between ges-
ture 1and gesture 2illustrate different aspects of coarticula-
tion. In the ﬁrst case, the coarticulatory effect is minimised
since the movement stops at the transition. In the second case,
coarticulatory effects intervene since the movement is not al-
lowed to stop and fused boundaries between segments must
Real-Time Inference of Coarticulated Gestures
We analyze the coarticulation between the two gestures
through the two probabilistic models mentioned.
3by Michael and Max Egger http://www.anyma.ch/2009/
a b c d
Figure 1: The gestures used in the experiment and several instances of their sequences. The two basic gestures are two-
dimensional trajectories representing the symbols V (a) and O (b). Figures (c) and (d) respectively represent a sequence of
gestures 1 and 2, either with a pause between, or rapidly without break.
The ﬁrst model is an adaptive template-based following sys-
tem . In this case we learn ﬁrst the whole sequence of
gestures 1and 2without coarticulation (i.e. performed with
a short stop between them) as the template (Figure 2a-left).
Then, we observe how the adaptive model can match the coar-
ticulated sequence (Figure 2a-middle). The Figure 2a also
shows the alignment that is computed between each gesture
segment, 1and 2(Figure 2a-middle and right). The adap-
tive following system can track the co-articulated ﬁgure, and
the transition between the two gestures appears clearly on the
alignment (Figure 2a-right).
Such an approach nevertheless requires to record, thus ‘learn’
the complete pair of gesture or at least their transition, simi-
larly to speech where diphones are considered.
Next, we consider how the coarticulation can be modeled
when a statistical model learns each gesture separately. We
use a hierarchical hidden Markov model that encodes each
gesture through a learning procedure . Gesture 1and 2are
learned by considering the 10 examples of the isolated per-
formance. A 10-state HMM is built from these gestures. The
analysis is performed online on the same chosen sequences.
Figure 2b shows the results of the real-time continuous ges-
ture recognition for the coarticulated sequences.
The HMM approach is a discrete-state model which allows
for a sharp transition between both gestures. Nevertheless,
we observed that the transition cannot always be deﬁned as a
unique point, as it is illustrated in Figure 2b-right.
Precisely, in the case of the HMM-like approach, the model is
able to consistently follow the ﬁrst gesture (the time progres-
sion evolves continuously from 0 to 1). However, the transi-
tion between the two gestures results in short-term recogni-
tion errors. This problem is typical of real-time continuous
gesture recognition, where the ambiguities of coarticulation
results in ‘jumps’ of the recognition until one gesture is re-
solved after the transition is complete.
DISCUSSION AND PERSPECTIVES
The example presented in this paper illustrated that both mod-
els can account for coarticulation in a ﬁrst approximation.
Importantly, boundaries are fuzzy and there is a need for an
interpolation (or extrapolation) strategy. Usually, constraints
on transition are imposed on the user in order to ensure better
recognition results as observed in the case where a pause is
respected between both gestures. Nevertheless, we advocate
here to improve movement modeling to better take into ac-
count coarticulation. In the ﬁrst model, interpolating would
mean interpolate between two templates by allowing cross-
gestural dependencies. The applications of method proposed
in computer animation (e.g. Gibet et al ) could be inter-
esting to evaluate in this context. In the second model, cross-
gestural dependencies could also be envisaged but it would
require examples of such dependencies in order to capture
It can indeed be argue that large database containing several
ways of performing coarticulated gestures would improve
recognition even in the presence of coarticulation. However,
our point here is different: coarticulation intrinsically con-
tains important information that we should be able to char-
acterize per se in our computational models, and exploit in
the interaction. In particular, coarticulation can inform on the
degree of expertise in skilled movements, as usually found in
music and dance. Moreover, it would be beneﬁcial to relate
computational models to the notion of chunking . Coar-
ticulation typically occurs within a chunk, and segmentation
marks could be detected between chunks.
It is important to note that the example we provide in 2D is
only representative of the research questions we propose. The
use of 3D trajectory and other movement modalities such as
acceleration data would pose additional issues. The gener-
alisation of the methodology to a 3D movement remains an
important goal of this research.
As a perspective, we propose to take into account the follow-
ing points for computational movement models:
•Movement anticipation implies to take into account the in-
ﬂuence of precedent segments for describing forthcoming
Time (normalised samples)
Time (normalised samples)
Time (normalised samples)
(a) Tracking test with the template-based method using particle ﬁltering. The blue ﬁgure is the template performed with a
pause (V – Stop – 0) and the red ﬁgure shows the performed coarticulated gesture (no pause between Vand 0). The central
ﬁgure shows the spatial alignment between the two ﬁgures and the right ﬁgure shows the temporal alignment where the
transition is clearly visible.
(b) Recognition test with the Hierarchical HMM. The left ﬁgure shows the coarticulated
gestures (no pause between Vand 0). The color corresponds to the recognition results:
blue is for Vand red for 0. The right ﬁgure shows the recognition over time: the blue
curve shows the time progression (as decoded by the model) during the V, followed by
time progression during the 0. In both ﬁgures, the transition is ambiguous.
Figure 2: Results of (a) gesture tracking using template-based approach and (b) continuous gesture recognition using a Hierar-
chical Hidden Markov Model)
•Low-level segments are concatenated in longer phrases
through practice and learning, which produce movement
variations over time altering the initial shapes of the seg-
ments and containing expressive features. Thus, movement
features and vocabularies must be considered as evolving
over time and prone to user idiosyncrasies.
•As fused boundaries between segments might appear
through practice, segmentation should be adaptive. Hybrid
segmentation-interpolation approaches should be consid-
One approach is to consider a fully Bayesian movement rep-
resentation that would be take into account uncertainty cross-
gestures, various time scales and time courses in anticipation.
Other approaches are however possible and we hope that this
‘perspective’ paper could contribute to trigger important dis-
cussions on this topic.
This work was supported by the Marie Skodowska-Curie Ac-
tion of the European Union (H2020-MSCA-IF-2014, IF-GF,
grant agreement no. 659232), the Labex SMART (ANR-11-
LABX-65) supported by French state funds managed by the
ANR within the Investissements d’Avenir programme under
reference ANR-11-IDEX-0004-02, and by the Moving Sto-
ries research partnership (SSHRC Award 31639884).
1. Bhuyan, M., Kumar, D. A., MacDorman, K. F., and
Iwahori, Y. A novel set of features for continuous hand
gesture recognition. Journal on Multimodal User
Interfaces 8, 4 (2014), 333–343.
2. Bianco, T., Freour, V., Rasamimanana, N., Bevilaqua, F.,
e, R. On gestural variation and coarticulation
effects in sound control. In Lecture Notes in Computer
Science, vol. 5934. Springer, 2009, 134–145.
3. Bizzi, E., Mussa-Ivaldi, F. A., and Giszter, S.
Computations underlying the execution of movement: a
biological perspective. Science 253, 5017 (1991),
4. Bush, B. O. Modeling coarticulation in continuous
speech. PhD thesis, Oregon Health and Science
5. Caramiaux, B., Bevilacqua, F., and Tanaka, A. Beyond
recognition: using gesture variation for continuous
interaction. In CHI’13 Extended Abstracts on Human
Factors in Computing Systems, ACM (2013),
6. Caramiaux, B., Montecchio, N., Tanaka, A., and
Bevilacqua, F. Adaptive gesture recognition with
variation estimation for interactive systems. ACM
Transactions on Interactive Intelligent Systems (TiiS) 4,
4 (2015), 18.
7. Caramiaux, B., Wanderley, M. M., and Bevilacqua, F.
Segmenting and parsing instrumentalists’ gestures.
Journal of New Music Research 41, 1 (2012), 13–29.
8. Diedrichsen, J., and Kornysheva, K. Motor skill learning
between selection and execution. Trends in Cognitive
Sciences 19, 4 (2015), 227–233.
9. Franc¸oise, J., Roby-Brami, A., Riboud, N., and
Bevilacqua, F. Movement sequence analysis using
hidden markov models: A case study in tai chi
performance. In Proceedings of the 2nd International
Workshop on Movement and Computing (MOCO’15),
ACM (2015), 29–36.
10. Franc¸oise, J., Schnell, N., Borghesi, R., and Bevilacqua,
F. Probabilistic models for designing motion and sound
relationships. In Proceedings of the 2014 International
Conference on New Interfaces for Musical Expression
11. Gibet, S., Lebourque, T., and Marteau, P.-F. High-level
speciﬁcation and animation of communicative gestures.
Journal of Visual Languages & Computing 12, 6 (2001),
12. Godøy, R. I. Understanding coarticulation in musical
experience. In Sound, Music, and Motion. Springer,
13. Godøy, R. I., Jensenius, A., and Nymoen, K. Chunking
in music by coarticulation. Acta Acustica united with
Acustica 96, 4 (2010), 690–700.
14. Hardcastle, W. J., and Hewlett, N. Coarticulation:
Theory, data and techniques. Cambridge University
15. Iskarous, K., Mooshammer, C., Hoole, P., Recasens, D.,
Shadle, C. H., Saltzman, E., and Whalen, D. The
coarticulation/invariance scale: Mutual information as a
measure of coarticulation resistance, motor synergy, and
articulatory invariance. The Journal of the Acoustical
Society of America 134, 2 (2013), 1271–1282.
16. Kirsh, D. Embodied cognition and the magical future of
interaction design. ACM Transactions on
Computer-Human Interaction (TOCHI) 20, 1 (2013), 3.
17. Li, S., Wang, L., and Kong, D. Synthesis of sign
language co-articulation based on key frames.
Multimedia Tools and Applications 74, 6 (2015),
18. Morency, L.-P., Quattoni, A., and Darrell, T.
Latent-dynamic discriminative models for continuous
gesture recognition. In Computer Vision and Pattern
Recognition, 2007. CVPR’07. IEEE Conference on,
IEEE (2007), 1–8.
19. Murphy, K. P. Machine learning: a probabilistic
perspective. MIT press, 2012.
20. Palmer, C., and Deutsch, D. Music performance:
Movement and coordination. In The psychology of music
(Third Edition), D. Deutsch, Ed. Elsevier Press
Amsterdam, 2013, 405–422.
21. Pelachaud, C. Communication and coarticulation in
facial animation. PhD thesis, University of
22. Rasamimanana, N. H., and Bevilacqua, F. Effort-based
analysis of bowing movements: evidence of anticipation
effects. The Journal of New Music Research 37, 4
(2009), 339 – 351.
om, D., Flanagan, J. R., and Johansson, R. S. Skill
learning involves optimizing the linking of action
phases. Journal of neurophysiology 110, 6 (2013),
24. Schnell, N., R ¨
obel, A., Schwarz, D., Peeters, G., and
Borghesi, R. MuBu & Friends - Assembling Tools for
Content Based Real-Time Interactive Audio Processing
in Max/MSP. In Proceedings of the International
Computer Music Conference (ICMC) (2009).
25. Schultz, T., and Wand, M. Modeling coarticulation in
emg-based continuous speech recognition. Speech
Communication 52, 4 (2010), 341–353.
26. Segouat, J. A study of sign language coarticulation.
SIGACCESS Access. Comput., 93 (Jan. 2009), 31–38.
27. Shah, A., Barto, A. G., and Fagg, A. H. A dual process
account of coarticulation in motor skill acquisition.
Journal of motor behavior 45, 6 (2013), 531–549.
28. Song, Y., Demirdjian, D., and Davis, R. Continuous
body and hand gesture recognition for natural
human-computer interaction. ACM Transactions on
Interactive Intelligent Systems (TiiS) 2, 1 (2012), 5.
29. Sosnik, R., Hauptmann, B., Karni, A., and Flash, T.
When practice leads to co-articulation: the evolution of
geometrically deﬁned movement primitives.
Experimental Brain Research 156, 4 (2004), 422–438.
30. Viswanathan, N., Magnuson, J. S., and Fowler, C. A.
Information for coarticulation: Static signal properties
or formant dynamics? Journal of Experimental
Psychology: Human Perception and Performance 40, 3
31. Wobbrock, J., Wilson, A., and Li, Y. Gestures without
libraries, toolkits or training: a $1 recognizer for user
interface prototypes. In Proceedings of the 20th annual
ACM symposium on User interface software and
technology, ACM (2007), 159–168.
32. Yang, R., and Sarkar, S. Detecting coarticulation in sign
language using conditional random ﬁelds. In Pattern
Recognition, 2006. ICPR 2006. 18th International
Conference on, vol. 2, IEEE (2006), 108–112.