Content uploaded by David Doukhan
Author content
All content in this area was uploaded by David Doukhan on Sep 19, 2024
Content may be subject to copyright.
Detecting the terminality of speech-turn boundary for spoken interactions in
French TV and Radio content
R´
emi Uro1,2, Marie Tahon3, David Doukhan1, Antoine Laurent3, Albert Rilliard2
1French National Institute of Audiovisual (INA), Paris, France. 2Universit´
e Paris Saclay, CNRS,
LISN, Orsay, France. 3LIUM, Le Mans Universit´
e, France
ruro@ina.fr, marie.tahon@univ-lemans.fr, ddoukhan@ina.fr,
antoine.laurent@univ-lemans.fr, albert.rilliard@lisn.fr
Abstract
Transition Relevance Places are defined as the end of an
utterance where the interlocutor may take the floor without in-
terrupting the current speaker –i.e., a place where the turn is
terminal. Analyzing turn terminality is useful to study the dy-
namic of turn-taking in spontaneous conversations. This paper
presents an automatic classification of spoken utterances as Ter-
minal or Non-Terminal in multi-speaker settings. We compared
audio, text, and fusions of both approaches on a French corpus
of TV and Radio extracts annotated with turn-terminality infor-
mation at each speaker change. Our models are based on pre-
trained self-supervised representations. We report results for
different fusion strategies and varying context sizes. This study
also questions the problem of performance variability by ana-
lyzing the differences in results for multiple training runs with
random initialization. The measured accuracy would allow the
use of these models for large-scale analysis of turn-taking.
Index Terms: Spoken interaction, Media, TV, Radio,
Transition-Relevance Places, Turn Taking, Interruption
1. Introduction
Turn-taking dynamics [1] and its relations to interruptions
are important factors in describing broadcast interactions, as
they hint at perceived power and dominance in a conversation
[2, 3]. [4] introduced the notion of Transition Relevance Places
(TRPs), which occur at the end of turn-construction units and
allow for smooth transitions between speakers. TRPs are im-
portant because speakers can anticipate their occurrences in or-
der to plan a potential turn-taking, optimizing the floor trans-
fer offset [5]. Automatically detecting TRPs would allow for
large-scale analysis of complex interruptions-related phenom-
ena. As for other human science objects, allowing large datasets
describing turn-tacking dynamics to be produced and analyzed
is an important outcome for society. While [6] focuses on over-
lapping speech segments to study interruptions following a clas-
sification schema defined by [7] for French political interviews,
interruptions may also occur without overlapping speech [8]. A
better understanding of turn-taking also has implications for the
development of human-machine interactions [9, 10]. [11] has
annotated TRPs in 8 hours of Slovak TV discussions and reports
a binary classification accuracy of 94.4% with an ensemble
model using fundamental frequency (F0) and intensity curves
on chunks of 1.5 s. [12, 13] propose an LSTM-based architec-
ture using acoustic and linguistic features on English sponta-
neous dialogues to predict whether a speaker change would oc-
cur in the three following seconds. [14] presents a Transformer-
based approach to the same task based on textual information
only. [9] proposed a method for turn-taking prediction in spo-
ken dialog systems. They report substantial inter-rater agree-
ments for annotating TRPs in different acted dialogue scenarios.
The proposed LSTM architecture using acoustic and linguistic
features achieves binary accuracies between 79.3 and 89.5.
The work presented here focuses on media broadcasts, es-
pecially for shows proposing multispeaker interactions. This
choice was made to investigate spontaneous spoken dialogs and
to allow collaboration with sociologists analyzing roles and be-
havior in media. We present an automatic classification of spo-
ken utterances as Terminal (i.e., ending with a TRP) or Non-
Terminal (not ending with a TRP). The corpus was based on
French TV and radio content, specifically multi-party conversa-
tion representing spontaneous interactions with various levels of
control [15]. An existing corpus annotated with TRPs [16] was
used to train multimodal models based on the Wav2Vec2 [17]
and FlauBERT [18] pre-trained models. We report results com-
paring different fusion strategies and context sizes, evaluate the
relative role of the different information modalities (lexical vs.
acoustic) with respect to the length of the segments used for the
inference, and discuss the relevance of these findings for spoken
interaction modeling.
2. Method
2.1. Data
This study is based on an annotated corpus of multi-party in-
teractions composed of French TV and Radio broadcasts from
1998 to 2015 [16] from ALLIES corpus [19]. Audio chunks
corresponding to zones of turn changes were extracted from the
complete show to encompass (1) the last speech segment1be-
fore the turn change and (2) the first segment after the change.
The segment (1) comes before the turn change and does not
include potential overlapped speech, while segment (2) may
contain an overlap. These audio chunks were annotated re-
garding the terminality of the initial speaker’s turn (terminal
or not) and the second speaker’s turn-taking category (interrup-
tion, backchannel, or smooth change). One annotator paid for
this work (a student trained in linguistics) made the annotations.
Two additional persons also annotated a subset of 338 samples
to obtain an inter-rater agreement. A substantial agreement of
0.75 using Fleiss’ Kappa [21] was observed for the Terminality
annotation used in this paper. The work presented here focuses
solely on the parts (1), preceding the speaker change, to avoid
relying on information linked to the second speaker’s interven-
tion. This was decided because we aim to produce a model that
classifies speech turn in relation to turn-taking management and
the anticipation of a given speaker floor taking [5], thus with-
out knowing what happened later during the dialogue. We se-
1defined by the Allies terminology as “sequences containing com-
plete words which are syntactically and semantically coherent” [20]
Interspeech 2024
1-5 September 2024, Kos, Greece
3560 10.21437/Interspeech.2024-1163
lected the segment (1) occurring before the speaker changes,
which constitute the ”samples” that are used in the remain-
ing of this study. These samples amounted to a total of 1954
speaker changes annotated with TRP information with 128 dif-
ferent speakers. Table 1 shows the number of samples and their
duration characteristics for each Terminality class.
Nb ≤0.5 0.5<x≤1 1<x≤2>2
Term. 839 13% 27% 35% 25%
Non Term. 1115 11% 19% 23% 47%
Table 1: Number of samples and percentage of samples in dif-
ferent duration intervals (in seconds) for the Terminal and Non-
Terminal classes
Table 2 lists the names of the different shows used to select
the samplesand the number and total duration of the annotated
samples from each show.
Two settings for the chunks used as input of the model
were tested: (1) the variable-size samples defined above, that
comprises a complete ”segment” as defined in the Allies corpus
(thereafter ref ); or (2) a fixed size chunk that correspond to the
three seconds before the turn taking event occurring at the end
of segment 1, and that could have been obtained using an auto-
matic diarization (thereafter 3s; these may include other speak-
ers turns at the beginning, or incomplete turn). This was done
to evaluate the possibility to apply a fully automatic approach.
For the fixed-sized chunks, other sizes have been tested in
preliminary evaluations (2 and 5 seconds) with similar results.
Three-second chunks were kept specifically to investigate if the
model would work even if provided with parts of previous turns.
2.2. Data preprocessing
For each sample, we extracted the relevant audio waveform, ei-
ther the annotated segment for ref or the last 3 seconds of the
recordings before the speaker change for 3s. All samples were
then automatically transcribed using Whisper [22]. To evaluate
the model in fully automatic processing pipeline conditions, we
kept the Whisper transcriptions as is, even if easily detectable
errors occurred (e.g., ”Subtitles generated by. . . ”). Manual
transcriptions for the ref setting were also available from the
Allies reference transcriptions and are used for comparison.
2.3. Model architectures
A total of 5 prediction models are proposed, operating on raw
audio and/or textual transcriptions. The models are based on the
pre-trained self-supervised models wav2vec2-base [17] for
the audio and flaubert base [18] for the text. They both
output a classification token (CLS) of size 768 from variable-
size audio or text inputs. Text-only (TO) and Audio-only (AO)
consist of three linear layers after extracting the CLS token of
the pre-trained model (see Figure 1). Three additional fusion
processes inspired by [23] are defined: Early Fusion (EF), with
one linear layer after each pre-trained model before concatena-
tion and another three linear layers (see Figure 2); Late Fusion
(LF) with one linear layer on the concatenation of the TO and
AO outputs; and Average Fusion (AF) which takes the average
of the logits of TO and AO. A dropout of 0.30 was applied be-
tween each linear layer during training.
Show Samples Dur. (s)
BFMStory (BS) 200 853.34
CaVousRegarde (CR) 269 969.84
CultureEtVous (CV) 7 8.69
DEBATE (D) 128 266.22
EntreLesLignes(EL) 307 1179.53
LaPlaceDuVillage (PV) 770 1316.76
PileEtFace (PF) 150 573.33
PlaneteShowbiz (PS) 10 13.94
TopQuestions (TQ) 5 21.88
fm (FM) 108 254.25
Table 2: Number samples from each show, and their total dura-
tion
Linear (520)
Linear (256)
Linear (2)
Pretrained (768)
Figure 1: Single modality model architecture
3. Results
For better convenience, the model’s input used during training
or at inference time will be referred to using the following sym-
bols: ref auto: reference speaker segmentation and their au-
tomatic speech transcription; ref man: reference speaker seg-
mentation and their manual speech transcription; 3s auto fixed
3-second duration excerpts and their automatic speech tran-
scription.
3.1. Evaluation protocol
Models were trained and cross-tested for each train and in-
ference input combination (ref auto,ref man and 3s auto), re-
sulting in nine evaluation configurations for each architecture.
Each model was evaluated using K-Fold grouped by show from
which the examples were extracted. This was done to prevent
having samples of the same show in the training and testing
sets and to allow for a comparison of the performance degrada-
tion associated with materials of different natures (e.g., different
types of shows). A random split of 30% of the training set was
used as a validation set. A patience of five epochs was used on
validation accuracy. This process was run 10 times to estimate
the variability due to random initialization, allowing for more
reliable accuracy measures (mean and confidence intervals over
the 10 runs are given). The training was done on an RTX 4090
GPU, and it took about 80 hours to train the 1500 models for
the whole experiment. In order to investigate the significance
of accuracy variation and the impact of the different settings, a
linear mixed model was fit on the accuracy of these model ar-
chitectures across training and test sets, with the samples nested
3561
Linear (768) Linear (768)
FlauBERT (768)
Linear (520)
Linear (256)
Linear (2)
Sigmoid
Wav2Vec2 (768)
Figure 2: Early fusion model architecture
in the show as random factors, as in equation 1.
accuracy ∼model∗(test+train)+(1|show/sample)(1)
where the accuracy is explained by the model architecture and
its interactions with the test and train settings as fixed factors,
and the sample nested in showas random factors. The model
was fitted using R’s lme4 library [24, 25]. The mean and con-
fidence interval fitted by the models are plotted in Figure 3. A
post hoc comparison of the models’ accuracies (using Bonfer-
roni correction) for each level of test and train settings was done.
3.2. Variation across models
All models achieved an average accuracy score above 0.85 as
can be seen on Figure 3. This indicates that the proposed ap-
proach reached coherent and qualitative results. Figure 3 shows
the mean accuracies with their confidence intervals for each
combination of model architecture, training set, and testing set,
as estimated by the linear mixed model across all random initial-
ization runs. We observed that the Text Only approach performs
significantly worse than the other models. Then comes the Late
and Average Fusion models, which significantly outperform the
TO approach and achieved comparable performances (in terms
of accuracy) but are significantly outperformed by the Early Fu-
sion and Audio Only approaches. AO and EF achieve compara-
ble performances, hinting that the audio signal carries an impor-
tant share of information for this task, including non-linguistic
cues. These trends are similar for each training and testing set-
ting. The training settings do not seem to have much impact, all
models achieved similar accuracies across training configura-
tions, with the ones trained on the ref man performing slightly
better.
Tables 3 and 4 report the mean accuracy for each show, re-
spectively, with the ref man and 3s models for the different test
settings. The EF model performs slightly better than the AO
for most shows. Meanwhile, considering the confidence inter-
vals obtained from the ten random runs, the difference does not
reach significance. The worst accuracies for both training set-
tings are observed on the PlanetShowbiz show, which contains
fewer shows (10) and displays an interaction style that differs
from the others, with presenters often reading prepared scripts.
Train settings: 3s_auto
Train settings: ref_man
Train settings: ref_auto
AO AF EF LF TO
0.75
0.80
0.85
0.90
0.95
1.00
0.75
0.80
0.85
0.90
0.95
1.00
0.75
0.80
0.85
0.90
0.95
1.00
Models' architecture
Accuracy
Test settings
ref_auto
ref_man
3s_auto
Figure 3: Mean accuracies for each model’s Test settings (lines)
with their confidence intervals, for each Train setting (subplots),
for the different model architectures (x-axis).
Using the reference segmentation leads to better results than us-
ing a fixed size of three seconds, but a fully automatic approach
still reaches accuracies over 90%.
Figure 4 represents the mean accuracy for each model ar-
chitecture depending on the training setting and the duration of
the speech segment in the training sample. We observe that tex-
tual information on shorter segments is less reliable than the
audio signal, but tends to allow similar performances on longer
segments. This is especially true with the automatic settings, as
very short segments may not include enough data for a reliable
transcription.
4. Discussion & conclusions
We compared different models and data settings to classify
the terminality status of speech samples extracted from spon-
taneous interactions in media archives. While the proposed
3562
Show AO EF
ref 3s ref ref man 3s
(BS) 97.86 96.21 98.14 98.21 95.79
(CR) 96.49 94.74 95.27 96.02 93.73
(CV) 97.96 87.76 95.92 95.92 89.80
(D) 96.54 94.42 96.09 96.54 94.20
(EL) 97.16 94.23 93.62 97.16 88.83
(PV) 91.41 87.25 93.75 95.84 90.95
(PF) 98.76 97.43 95.90 99.33 93.81
(PS) 84.29 77.14 82.86 84.29 74.29
(TQ) 91.43 91.43 100.00 100.00 100.00
(FM) 96.16 94.18 96.16 95.77 95.24
Mean 94.80 91.48 94.77 95.90 91.66
Table 3: Mean accuracies for each show with the AO and EF
model trained in the ref man setting
Show AO EF
ref 3s ref ref man 3s
(BS) 97.86 97.64 98.29 98.29 97.50
(CR) 96.49 96.23 96.81 96.81 95.75
(CV) 93.88 93.88 93.88 93.88 91.84
(D) 96.76 97.66 95.87 95.76 96.76
(EL) 97.86 95.86 97.95 97.91 96.04
(PV) 88.09 85.77 96.07 96.38 93.65
(PF) 98.86 98.86 98.48 98.95 97.81
(PS) 87.14 81.43 87.14 87.14 81.43
(TQ) 94.29 88.57 94.29 94.29 91.43
(FM) 96.16 94.58 96.30 96.30 95.24
Mean 94.74 93.05 95.50 95.56 93.75
Table 4: Mean accuracies for each show with the AO and EF
model trained in the 3s setting
approaches achieved highly encouraging performances, these
models are based on a corpus of about 2000 audio samples
from 10 different shows with limited types of interactions and
have not yet been tested on other corpora. One important take-
away of this study is that the Wav2Vec2 representation alone
seems sufficient for classifying Transition Relevance Places, as
Early Fusion and Audio Only models achieve comparable per-
formances. However, the AO model is smaller and does not
require transcription as a preprocessing step, making it less
resource-demanding and thus more easily usable in real-life set-
tings. It also potentially has fewer language dependencies than
a model (partially) based on text, albeit we didn’t test this. The
Text Only models also achieve satisfying performances, espe-
cially on longer speech turns, showing that lexical information
is relevant for this task but, as pointed out by [14], more work
on language modeling, specifically focused on the varying lev-
els of boundaries (e.g., [26]), would be important. Compar-
ing different settings for data preprocessing (manual segmenta-
tion vs. fixed-size window; manual vs. automatic transcription)
shows that the proposed models are fairly invariant to the pre-
processing setting, hinting that they could be used in a fully au-
tomatic setting. Applying automatic diarization may also help
to reduce the gap between manual and automatic settings at in-
ference. Testing the proposed models on other languages and
interaction types could prove useful in better understanding the
differences in turn-taking dynamics depending on context. The
performances that the proposed models achieved are coherent
with the most similar works in the literature [11] which is based
on prosodic features alone. We believe rendering our dataset
0.0 0.2 0.4 0.6 0.8 1.0
Duration (s)
0.0
0.2
0.4
0.6
0.8
1.0
Accuracy
0246810
0.7
0.8
0.9
1.0 ref_auto
0246810
0.7
0.8
0.9
1.0 ref_man
0246810
0.7
0.8
0.9
1.0 3s_auto
TO
AO
EF
LF
AF
Figure 4: Mean accuracies as a function of the duration of the
testing sample for the different three training settings.
available will improve evaluation homogeneity among similar
work. We think the notion of terminal speech turn is important
for better-describing interactions, and the results we obtained
here encourage us to assert that it is a notion that can be auto-
matically predicted from spontaneous speech. This work can
have applications for large-scale semi-automatic media analy-
sis. The annotation of interruptions in the media is a subjective
task, with multiple studies not necessarily obtaining the same
conclusions [27, 28]. The proposed system would allow for
a reproducible method for automatically highlighting places of
interest in a long document and thus enable researchers to fo-
cus on the speaker, specifically changes that happen outside of a
detected TRP on French TV and Radio content. This corpus is
also annotated with backchannels and interruption information,
future works will focus on the classification of turn-taking with
regard to these phenomena.
5. Availability
Code and datasets used in this paper are available at https:
//github.com/ina-foss/termClassif.
3563
6. Acknowledgements
This work has been partially funded by the French National Re-
search Agency and the German DFG (GEM project - ANR-19-
CE38-0012 and CLD 2025 - ANR-19-CE38-0015).
7. References
[1] E. Couper-Kuhlen, English Speech Rhythm, ser. Pragmatics Be-
yond New Series. Amsterdam, Netherlands: John Benjamins
Publishing, Apr. 1993.
[2] D. H. Zimmerman and C. West, “Sex roles, interruptions and si-
lences in conversation,” 1996.
[3] G. W. Beattie, “Turn-taking and interruption in political
interviews: Margaret thatcher and jim callaghan compared
and contrasted,” Semiotica, vol. 39, no. 1–2, 1982. [On-
line]. Available: https://www.degruyter.com/document/doi/10.
1515/semi.1982.39.1-2.93/html
[4] H. Sacks, E. A. Schegloff, and G. Jefferson, “A simplest system-
atics for the organization of turn-taking for conversation,” Lan-
guage, vol. 50, no. 4, p. 696, Dec. 1974.
[5] S. C. Levinson, “Turn-taking in human communication – origins
and implications for language processing,” Trends in Cognitive
Sciences, vol. 20, no. 1, p. 6–14, Jan. 2016.
[6] M. Lebourdais, M. Tahon, A. Laurent, S. Meignier, and
A. Larcher, “Overlaps and gender analysis in the context of broad-
cast media.” LREC 2022, 2022, p. 7.
[7] M. Adda-Decker, C. Barras, G. Adda, P. Paroubek, P. Boula de
Mare¨
uil, and B. Habert, “Annotation and analysis of overlapping
speech in political interviews,” in LREC 2008, Marrakech, Mo-
rocco, 2008, pp. 1–7.
[8] S. C. Levinson, Pragmatics. Cambridge University Press, 1983.
[9] K. Hara, K. Inoue, K. Takanashi, and T. Kawahara, “Turn-taking
prediction based on detection of transition relevance place,” in
Interspeech 2019. ISCA, Sep. 2019, p. 4170–4174. [Online].
Available: http://www.isca-speech.org/archive/Interspeech 2019/
abstracts/1537.html
[10] G. Skantze, “Turn-taking in conversational systems and human-
robot interaction: A review,” Computer Speech Language,
vol. 67, p. 101178, May 2021.
[11] S. Ondas, M. Pleva, and S. Bacikova, “Transition-relevance places
machine learning-based detection in dialogue interactions,” Elek-
tronika ir Elektrotechnika, vol. 29, no. 3, p. 48–54, 2023.
[12] G. Skantze, “Towards a general, continuous model of turn-taking
in spoken dialogue using lstm recurrent neural networks,”
in Proceedings of the 18th Annual SIGdial Meeting on
Discourse and Dialogue. Saarbr¨
ucken, Germany: Association
for Computational Linguistics, 2017, p. 220–230. [Online].
Available: http://aclweb.org/anthology/W17-5527
[13] M. Roddy, G. Skantze, and N. Harte, “Multimodal continuous
turn-taking prediction using multiscale rnns,” in Proceedings
of the 20th ACM International Conference on Multimodal
Interaction, Oct. 2018, p. 186–190, arXiv:1808.10785 [cs].
[Online]. Available: http://arxiv.org/abs/1808.10785
[14] E. Ekstedt and G. Skantze, “Turngpt: a transformer-based
language model for predicting turn-taking in spoken dialog,”
in Findings of the Association for Computational Linguistics:
EMNLP 2020, 2020, p. 2981–2990, arXiv:2010.10874 [cs].
[Online]. Available: http://arxiv.org/abs/2010.10874
[15] P. Wagner, J. Trouvain, and F. Zimmerer, “In defense of stylistic
diversity in speech research,” Journal of Phonetics, vol. 48, p.
1–12, 2015.
[16] R. Uro, M. Tahon, J. Wottawa, D. Doukhan, A. Rilliard,
and A. Laurent, “Annotation of transition-relevance places and
interruptions for the description of turn-taking in conversations
in French media content,” in Proceedings of the 2024
Joint International Conference on Computational Linguistics,
Language Resources and Evaluation (LREC-COLING 2024),
N. Calzolari, M.-Y. Kan, V. Hoste, A. Lenci, S. Sakti, and N. Xue,
Eds. Torino, Italia: ELRA and ICCL, May 2024, pp. 1225–1232.
[Online]. Available: https://aclanthology.org/2024.lrec-main.110
[17] A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0:
A framework for self-supervised learning of speech representa-
tions,” no. arXiv:2006.11477, Oct. 2020, arXiv:2006.11477 [cs,
eess]. [Online]. Available: http://arxiv.org/abs/2006.11477
[18] H. Le, L. Vial, J. Frej, V. Segonne, M. Coavoux, B. Lecouteux,
A. Allauzen, B. Crabb´
e, L. Besacier, and D. Schwab,
“FlauBERT: Unsupervised language model pre-training for
French,” in Proceedings of the Twelfth Language Resources
and Evaluation Conference, N. Calzolari, F. B´
echet, P. Blache,
K. Choukri, C. Cieri, T. Declerck, S. Goggi, H. Isahara,
B. Maegaard, J. Mariani, H. Mazo, A. Moreno, J. Odijk,
and S. Piperidis, Eds. Marseille, France: European Language
Resources Association, May 2020, pp. 2479–2490. [Online].
Available: https://aclanthology.org/2020.lrec-1.302
[19] M. Tahon, A. Larcher, M. Lebourdais, F. Bougares, A. Silnova,
and P. Gimeno, “Allies: a speech corpus for segmentation, speaker
diarization speech recognition and speaker change detection,” in
LREC/COLING, 2024.
[20] A. Larcher, A. Mehrish, M. Tahon, S. Meignier, J. Carrive,
D. Doukhan, O. Galibert, and N. Evans, “Speaker Embedding
For Diarization Of Broadcast Data In The ALLIES Challenge.”
ICASSP, 2021, pp. 5799–5803.
[21] J. L. Fleiss, “Measuring nominal scale agreement among many
raters.” Psychological Bulletin, vol. 76, no. 5, p. 378–382, 1971.
[22] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey,
and I. Sutskever, “Robust speech recognition via large-
scale weak supervision,” no. arXiv:2212.04356, Dec. 2022,
arXiv:2212.04356 [cs, eess]. [Online]. Available: http://arxiv.org/
abs/2212.04356
[23] M. Macary, M. Tahon, Y. Est `
eve, and D. Luzzati, “Acoustic
and linguistic representations for speech continuous emotion
recognition in call center conversations,” no. arXiv:2310.04481,
2023. [Online]. Available: https://arxiv.org/abs/2310.04481v1
[24] R Core Team, R: A Language and Environment for Statistical
Computing, R Foundation for Statistical Computing, Vienna,
Austria, 2023. [Online]. Available: https://www.R-project.org/
[25] D. Bates, M. M¨
achler, B. Bolker, and S. Walker, “Fitting linear
mixed-effects models using lme4,” Journal of Statistical Software,
vol. 67, no. 1, 2015.
[26] P. A. Barbosa and T. Raso, “Spontaneous speech segmenta-
tion: Functional and prosodic aspects with applications for au-
tomatic segmentation/a segmentac¸˜
ao da fala espontˆ
anea: aspectos
pros´
odicos, funcionais e aplicac¸ ˜
oes para a tecnologia,” Revista de
Estudos da Linguagem, vol. 26, no. 4, pp. 1361–1396, 2018.
[27] M. Sandr´
e, “Analyse d’un dysfonctionnement interactionnel –
l’interruption – dans le d´
ebat de l’entre-deux-tours de l’´
election
pr´
esidentielle de 2007,” Mots. Les langages du politique, no.
8989, p. 69–81, Mar. 2009.
[28] H. Constantin de Chanay and C. Kerbrat-Orecchioni, “Les in-
terruptions dans les d´
ebats m´
ediatiques: une strat ´
egie interac-
tionnelle,” Pratiques. Linguistique, litt ´
erature, didactique, no.
147–148147–148, p. 83–104, Dec. 2010.
3564
