ARTICULATORY BEHAVIOUR DURING DISFLUENCIES IN
Ivana Didirková1,2, Sébastien Le Maguer3, Fabrice Hirsch4, Dodji Gbedahou4
1EA 1569 TransCrit, Université Paris 8, France,
2CNRS & UMR 7018 Laboratoire de phonétique et phonologie, Université Paris 3, France
3ADAPT Centre, Sigmedia Lab, EE Engineering, Trinity College Dublin, Dublin, Ireland
4CNRS & UMR 5267 Praxiling, Université Paul-Valéry Montpellier 3, France
firstname.lastname@example.org, email@example.com, firstname.lastname@example.org, email@example.com
The aim of this study is to analyse articulatory
movements that occur during Stuttering-Like Dis-
ﬂuencies (SLD) and to propose a new classiﬁca-
tion of SLD based on supraglottic articulatory ges-
tures. To carry out this study, ElectroMagnetic Ar-
ticulography (EMA) data were collected within two
Persons Who Stutter (PWS) reading two texts. All
pathological disﬂuencies were identiﬁed in the pro-
duction of PWS categorized as blocks, repetitions
and prolongations. Results show four articulatory
patterns occurring during the SLD: Reiterations of
series of movements leading to sound(s) or syllable
repetitions, global maintain of the articulatory pos-
ture, anarchical movements and a combination of
above. While the ﬁrst category only concerns repeti-
tions, the three others can concern SLD categorized
as repetitions, prolongations or blocks.
Keywords: stuttering; stuttering-like disﬂuencies;
speech production; articulatory description.
Stuttering can be deﬁned as an alteration of speech
ﬂuency having negative implications on communic-
ation (). More precisely, this disorder is con-
sidered as a motor trouble that momentarily stops
speech ﬂow. Several types of stuttering are men-
tioned in literature: developmental stuttering start-
ing between age 3 and 7 and disappearing spontan-
eously, persistent stuttering beginning at the same
period but remaining present in adolescence and
adulthood; as well as acquired stuttering, gener-
ally due to a neurological accident (). Accord-
ing to , 5% of the worldwide population have
been concerned by this disorder but its prevalence
is at 1% since the rate of ‘spontaneous’ remissions
in children is evaluated at 80%. If the origins of
stuttering remain a challenge for researchers, recent
works allow formulating several hypotheses about
the aetiology of developmental and persistent stut-
tering. Indeed, the origins of this trouble should be
multi-factorial since different studies point out ge-
netic and neurological speciﬁcities in Persons Who
1.2. Phonetics of stuttering
As mentioned above, stuttered speech is charac-
terized by the presence of disﬂuencies that are
more frequent than in non-stuttered speech. These
Stuttering-Like Disﬂuencies (SLD) can be classi-
ﬁed mainly as blocks, prolongations and repetitions
(e.g. ) but other types of speech ﬂow alterations
can be found in PWS (see  for a literature review).
Moreover, SLD present several speciﬁcities com-
pared to non-pathological disﬂuencies. For ex-
ample, stuttering is one of the scarce disorders where
disﬂuencies can frequently split a syllable ().
In another study,  show that alterations of
speech ﬂow by PWS are generally accompanied by
audible tensions. The same research shows that con-
sonants can be prolonged in stuttering-like disﬂuen-
cies, while this is not the case in normal alterations
of speech ﬂow in French. Finally, they observe that
the duration of SLD is generally more important and
1.3. Physiological description of SLD
However, classiﬁcation of speech disﬂuencies as
non-pathological or pathological is not an easy task
given that several types of disﬂuencies are present
both in non-stuttered and stuttered speech. For in-
stance, sound prolongations, repetitions, as well as
silences, can also be observed in people who do not
stutter. This is the reason why physiological descrip-
tions are necessary to determine what distinguish
stuttering-like disﬂuencies and non pathological dis-
Concerning the respiratory level,  observe that
the respiratory movements during pre-phonatory
phases are different in PWS. Other studies dealing
with this topic have been carried out (e.g. , ).
The laryngeal level also presents some speciﬁcities.
Indeed,  observe an abnormal activity of the vo-
cal folds during stuttering-like disﬂuencies. There-
fore,  prefer to talk about myoclonic movements
(spasms) to describe the glottis functioning in PWS.
Curiously, The literature concerning the supra-
glottic level in subjects who stutter remains scarse
and often deal with ﬂuent speech produced by
PWS . Among studies dealing with SLD,  re-
veal a deﬁciency in the jaw-phonatory connection.
The speech motor behaviour in PWS would tend
to be less efﬁcient or even immature in the man-
agement of the coordination of different articulat-
ors. Thus,  supposes alterations of speech ﬂow
are due to a coarticulation disruption. More pre-
cisely,  estimates that, in a sequence, trans-
ition between the two sounds should be the con-
sequence of a disrupted antagonist muscles activ-
ity. This fault line would correspond to the moment
where stuttering-like disﬂuencies emerge.
1.4. Objective and hypothesis
As mentioned above, few studies have been carried
out on the way disﬂuencies are produced. Further-
more, most of these studies are based on extrapola-
tions made from acoustic data. However, while it is
possible to obtain many informations thanks to the
acoustic signal, EMA data allow a more direct ob-
Consequently, the aim of this study is to provide
a description of articulatory behaviour during SLD.
More precisely, our objective is to analyse articulat-
ory movements that occur during SLD and to pro-
pose a classiﬁcation based on supraglottic articu-
latory gestures. Our hypothesis is that the nomen-
clature generally used to describe disﬂuencies does
not reﬂect the articulatory behaviour. If a same per-
cept can be a result of different articulatory gestures
(), depending on speaker, phonetic environment,
etc., we postulate that the same articulatory patterns
could be at the origin of several perceptual types of
2.1. Data acquisition & participants
EMA data were collected by means of an electro-
magnetic articulograph Carstens AG501 3D at the
Lorraine Research Laboratory in Computer Science
and its Applications (LORIA, Nancy, France) with
a sampling rate of 250 Hz and an accuracy of 0.3
mm. All data were stocked in a .pos ﬁle and syn-
chronized with a sound recording (44.1 kHz, 16 bits,
.wav). 10 sensors (2x3 mm) per subject were used:
two were ﬁxed on the lips of each subject (1 in
the middle of the upper lip and another one in the
middle of the lower lip). 3 coils were situated on the
tongue of each subject; one on the tongue tip, one
on the tongue body and one on the tongue back. To
track the mandible’s movements, another sensor was
placed on the subjects’ jaw. The palate’s form was
indicated by means of a seventh coil. Other sensors
were used to control head’s movements.
Two PWS, one female and one male, aged re-
spectively 23 and 26, both native speakers of French
and Wolof, were recruited for this study. Participants
were recorded while reading the text of an Alphonse
Daudet’s novel, La chèvre de Monsieur Seguin (Mis-
ter Seguin’s goat), in French and an Aesop’s fable,
Le lion et le rat (The lion and the rat). These record-
ings took place in a soundproof room.
2.2. Data analysis
2.2.1. Acoustic and perceptual analysis
Data analysis rely on perceptual and acoustic iden-
tiﬁcation of stuttering-like disﬂuencies. First, three
persons (two of the authors and a speech therapist
specialized in stuttering) identiﬁed all SLD in the
production of PWS, based on perception and on the
speech signal, without classifying these disﬂuencies.
They then discussed cases where they did not reach
agreement. In order to conﬁrm their annotations and
identify the perceptual class of every disﬂuency, a
perception test has been carried out within the free-
ware Perceval (), based on .wav ﬁles extracted for
each SLD. Five naïve listeners were then asked to
categorize SLD as blocks, repetitions, prolongations
or combined disﬂuencies (Fleiss’ kappa: 0.752).
Authors discussed cases where naïve listeners did
not reach agreement. Speech alterations identiﬁed
as combined disﬂuencies were eliminated from fur-
ther study. Moreover, repetitions of diphones, syl-
lables, words and other sequences containing more
than one phone were excluded from our research in
order to minimize inﬂuence of coarticulation on our
After exclusion of combined disﬂuencies and dis-
ﬂuencies concerning more than one phone, 250 SLD
were obtained. Their distribution according to the
perceptual type of disﬂuency and according to the
subject can be found in the Table 1. Although both
subjects have a severe stuttering, the distribution of
disﬂuencies is not the same: whereas speaker F pro-
duces 89 disﬂuencies, mostly blocks and prolonga-
tions, 161 of disﬂuencies analysed in this paper were
produced by M. For the speaker M, repetitions are
the most present perceptual disﬂuency type. Due
to these idiosyncratic characteristics of SLD in our
speakers, only 55 disﬂuencies (22%) were blocks.
Other 39.2% of disﬂuencies were prolongations. Re-
petitions represented 38.8% of all analysed SLD. All
of SLD were spontaneous, e.g. no factors were ma-
nipulated to elicit these disﬂuencies.
Table 1: Distribution of stuttering-like disﬂuen-
cies according to their perceptual type and accord-
ing to the speaker
Repet. Prolong. Block Total
Female 17 34 38 89
Male 80 64 17 161
Total 97 98 55 250
2.2.2. Automatic articulatory analysis
We can assess if there was a movement during the
production by inspecting the articulatory dynamics.
To do so, we have deﬁned the following methodo-
logy. First, we consider t∈[0..T]the index of the
frame and Ctthe set of coils at frame t. From each
coil ct∈Ct, we compute the local velocity based
the central ﬁnite difference as deﬁned in  using
(1) ∆(ct) =
Then, a movement at frame t, is detected if the fol-
lowing criterion is validated:
with f(v,θ) = 1 if vis beyond θand 0 else.
In this study, we deﬁne θcat 30% of the aver-
age dynamic of the whole corpus for each coil c.
This large threshold allows us to be less sensitive
to a movement and therefore enhance non-activity
detection. Finally, we focus our analysis to the seg-
ments annotated as a disﬂuent production. For each
segment, we ignore the 10% ﬁrst and the 10% last
frames in order to avoid transition effect. From
the remaining ones, we compute the percentage of
frames considered in movement.
2.2.3. Manual articulatory analysis
After this classiﬁcation, the Visartico software ()
was used to visualize and analyse the vertical move-
ments (the z axis) of the upper and lower lip, the
tongue tip, the tongue body, the tongue back and
the mandible in segments that included the stuttered
phone and its preceding and subsequent phones.
3.1. Percentage of frames in movement by type of
As we can see in Table 2, even though a threshold
has been deﬁned to capture a maximum of non-
movement frames, there are still around 40% of
them considered as moving in average. Further-
more, the standard deviation shows an important
variability across the segments. Some segments are
even reaching 80% of movement.
Table 2: Average movement percentage per type
and per disﬂuency type. The standard deviation is
indicated in parenthesis.
Repetition 46.28 (13.15) 48.05 (12.91)
Prolong. 31.18 (13.43) 40.80 (9.60)
Block 42.26 (11.69) 47.18 (11.08)
If we observe what happens for each type of
SLD, it is possible to notice, in average, that repeti-
tions are the disﬂuencies produced with the greatest
amount of frames in movement in speakers F (aver-
age: 46.28%, SD: 13.15%) and M (average: 48.05,
SD: 12.91%). Blocks constitute the second type
of SLD where frames in movement are the most
present in F (average: 42.26%, SD: 11.69%) and
M (average: 47.18%, SD: 11.08%). Finally, less
frames are in movement in prolongations (average:
31.18%, SD: 13.43 in F; average: 40.8%, SD: 9.6 in
3.2. SLD and articulatory patterns
Four main categories of disﬂuencies have been re-
vealed by EMA data (a chi-squared goodness-of-
ﬁt test: χ2=187.28, df=3, p=.000): a) Reiterations
of series of movements leading to sound repeti-
tions (rep); b) Combination of a global maintain of
an articulatory posture and articulatory movements
(comb); c) Global maintain of the articulatory pos-
ture with or without an acoustic output and with or
without anticipation of the subsequent phone (no–
mov); d) Presence of articulatory movements with
or without inter-articulatory coupling (mov).
As shown in Figure 1, while the ﬁrst category
mostly concerns repetitions, the three others can
concern SLD categorized as repetitions, prolonga-
tions or blocks, showing that a same articulatory
pattern can be observed for the 3 types of disﬂuen-
cies (chi-squared test for independence: χ2=39.302,
df=6, p=.000, effect size: 0.560).
Figure 1: Proportions of disﬂuency types. The
area of each rectangle gives the proportion of the
perceptual type (width) and the articulatory pat-
block prolongation repetition
Perceptual type of disﬂuency
Art. pattern comb movno-movrep
3.3. Duration and SLD type
A linear model was ﬁt with articulatory pattern and
perceived disﬂuency as the independent variables
and length as the dependent variable. The model
was signiﬁcant: F=11.68 on 5 and 244 df, p=.000.
Our data indicate a clear preference of the combined
articulatory pattern to occur within the longest dis-
ﬂuencies. The duration decreases when the disﬂu-
ency is characterized by the presence of a movement
during whole disﬂuency. The shortest disﬂuencies
are those where we observe a global maintain of ar-
ticulatory posture and a repetition of an articulatory
movement. These effects are mostly prominent in
blocks as shown in Figure 2.
4. DISCUSSION AND CONCLUSION
To sum up, most of SLD are carried out with frames
in movement. If movements are generally observed,
it is possible to note that their ‘efﬁciency’ is vari-
able: indeed, if some of them are audible during
prolongations and repetitions for instance, others are
ineffective since they are inaudible, as in blocks. It
Figure 2: Types of disﬂuencies and their length.
block prolongation repetition
Perceptual type of disﬂuency
Disﬂ. length in s
Art. pattern comb movno-movrep
is important to highlight that movements’ efﬁciency
during SLD can be due a) to the degree of constric-
tion between the different articulators and b) to the
respiratory and/or the laryngeal level. Indeed, if air
pressure and/or vocal folds conﬁguration are not ad-
apted, the acoustic output will be absent.
Moreover, several types of articulatory patterns
have been observed during disﬂuencies. These pat-
terns can be divided in two categories: those which
are carried out with slight vertical movements or an
immobilization of most articulators, and those pro-
duced with movements. Among the last category
cited, there are SLD presenting inter-articulators
coupling and SLD where articulators move inde-
pendently of each other. These different patterns are
present for blocks, prolongations and repetitions. In
other terms, it means that a same type of disﬂuencies
can be produced in different ways. This allows to
draw a parallel between disﬂuencies and the Quan-
tal theory ; This theory supposes that a same per-
cept can be the results of several different articulat-
ors’ positions. As for disﬂuencies, a same disﬂuency
can be the result of different conﬁgurations.
Concerning articulatory patterns, it has been
shown that the longest disﬂuencies are carried out
with a combination of several articulatory conﬁgur-
ations. This result suggests that more there are dif-
ferent articulatory patterns during a disﬂuency and
longer the disﬂuency will be. Consequently, making
an effort seems to be ineffective in PWS when they
have to overcome a disﬂuency.
Our study reports articulatory patterns seen in
their totality (all articulators taken together). Thus,
it seems necessary to investigate the contribution
of each articulator to observed patterns. Our res-
ult should ﬁnally be compared to an articulatory de-
scription of disﬂuencies of non-stuttering speakers.
Finally, we think that this research should be car-
ried out in a longitudinal approach in order to verify
how articulatory patterns progress during a speech
therapy, showing that articulatory data should be
used more frequently during a stuttering reeducation
to note patient’s evolution.
This research was partly supported by « MoSpeeDi.
Motor Speech Disorders : characterizing phon-
etic speech planning and motor speech program-
ming/execution and their impairments », subside
CRSII5_173711/1Sinergia du Fond National Suisse
de la Recherche Scientiﬁque
This research was partly supported by the French
Agence Nationale de la Recherche and by the
Caisse nationale de solidarité pour l’autonomie un-
der Grant No. ANR-18-CE36-0008 (Project BE-
NEPHIDIRE, PI: Fabrice Hirsch).
This research was also partly supported by the
Irish Research Council (IRC) and by the ADAPT
Centre. The ADAPT Centre for Digital Con-
tent Technology is funded under the SFI Research
Centres Programme (Grant 13/RC/2106) and isco-
funded under the European Regional Development
 André, C., Ghio, A., Cavé, C., Teston, B. 2003.
PERCEVAL: a Computer-Driven System for Ex-
perimentation on Auditory and Visual Percep-
tion. International Congress of Phonetic Sciences
(ICPhS) Barcelona, Spain. UAB 1421–1424.
 Conture, E. G., Schwartz, H. D., Brewer, D. W.
1985. Laryngeal behavior during stuttering: A fur-
ther study. Journal of Speech, Language, and Hear-
ing Research 28(2), 233–240.
 Didirkova, I. 2016. Parole, langues et disﬂu-
ences: une étude linguistique et phonétique du
bégaiement. PhD thesis Université Paul Valéry-
 Didirkova, I., Fauth, C., Hirsch, F., Luxardo, G.,
Diwersy, S. 2016. Disﬂuences normales vs. disﬂu-
ences sévères: une étude acoustique. JEP-Journées
d’Etudes sur la Parole volume 1 191–199.
 Heyde, C. J., Scobbie, J. M., Lickley, R., Drake,
E. K. 2016. How ﬂuent is the ﬂuent speech of
people who stutter? a new approach to measuring
kinematics with ultrasound. Clinical linguistics &
phonetics 30(3-5), 292–312.
 Loucks, T. M., Luc, F., Sasisekaran, J. 2007. Jaw-
phonatory coordination in chronic developmental
stuttering. Journal of Communication Disorders
 Monfrais-Pfauwadel, M.-C. 2014. Bégaiement, bé-
gaiements. Un manuel clinique et thérapeutique.
 Monfrais-Pfauwadel, M.-C., Tromelin, O., Mou-
gin, A.-L., Ormezzano, Y. 2005. Utilisa-
tion des explorations multimédia synchrones dans
l’objectivation des événements laryngés lors des
bégayages. Revue de laryngologie, d’otologie et de
rhinologie 126(5), 341–345.
 Ouni, S., Mangeonjean, L., Steiner, I. Sept. 2012.
VisArtico: a visualization tool for articulatory
data. 13th Annual Conference of the International
Speech Communication Association - InterSpeech
2012 Portland, OR, United States.
 Peters, H. F., Boves, L. 1988. Coordination of aero-
dynamic and phonatory processes in ﬂuent speech
utterances of stutterers. Journal of Speech, Lan-
guage, and Hearing Research 31(3), 352–361.
 Piérart, B. 2011. Les bégaiements de l’adulte
volume 5. Editions Mardaga.
 Stevens, K. N., Hanson, H. M. 2012. Chapter 12:
Articulatory–acoustic relations as the basis of dis-
tinctive contrasts. The Handbook of Phonetic Sci-
ences 116, 424.
 Watson, B. C., Alfonso, P. J. 1987. Coordination of
prephonatory events in mild and severe stutterers.
In: Speech Motor Dynamics in Stuttering. Springer
 Wingate, M. E. 2012. The structure of stuttering:
A psycholinguistic analysis. Springer Science &
 Yairi, E., Ambrose, N. 2013. Epidemiology of stut-
tering: 21st century advances. Journal of ﬂuency
disorders 38(2), 66–87.
 Young, S., Evermann, G., Gales, M., Hain, T., Ker-
shaw, D., Liu, X., Moore, G., Odell, J., Ollason, D.,
Povey, D., others, 2006. The HTK book (v3. 4).
 Zellner, B. 1992. Le bé bégayage et euh...
l’hésitation en français spontané. JEP volume 19
 Zocchi, L., Estenne, M., Johnston, S., Del Ferro,
L., E. Ward, M., T. Macklem, P. 07 1990. Respirat-
ory muscle incoordination in stuttering speech. The
American review of respiratory disease 141, 1510–