Publications

Abstract

Biological structures are defined by rigid elements, such as bones, and elastic elements, like muscles and membranes. Computer vision advances have enabled automatic tracking of moving animal skeletal poses. Such developments provide insights into complex time-varying dynamics of biological motion. Conversely, the elastic soft-tissues of organisms, like the nose of elephant seals, or the buccal sac of frogs, are poorly studied and no computer vision methods have been proposed. This leaves major gaps in different areas of biology. In primatology, most critically, the function of air sacs is widely debated; many open questions on the role of air sacs in the evolution of animal communication, including human speech, remain unanswered. To support the dynamic study of soft-tissue structures, we present a toolkit for the automated tracking of semi-circular elastic structures in biological video data. The toolkit contains unsupervised computer vision tools (using Hough transform) and supervised deep learning (by adapting DeepLabCut) methodology to track inflation of laryngeal air sacs or other biological spherical objects (e.g., gular cavities). Confirming the value of elastic kinematic analysis, we show that air sac inflation correlates with acoustic markers that likely inform about body size. Finally, we present a pre-processed audiovisual-kinematic dataset of 7+ hours of closeup audiovisual recordings of siamang (Symphalangus syndactylus) singing. This toolkit (https://github.com/WimPouw/AirSacTracker) aims to revitalize the study of non-skeletal morphological structures across multiple species.

Abstract

Conversation requires a substantial amount of coordination between dialogue participants, from managing turn taking to negotiating mutual understanding. Part of this coordination effort surfaces as the reuse of linguistic behaviour across speakers, a process often referred to as alignment. While the presence of linguistic alignment is well documented in the literature, several questions remain open, including the extent to which patterns of reuse across speakers have an impact on the emergence of labelling conventions for novel referents. In this study, we put forward a methodology for automatically detecting shared lemmatised constructions---expressions with a common lexical core used by both speakers within a dialogue---and apply it to a referential communication corpus where participants aim to identify novel objects for which no established labels exist. Our analyses uncover the usage patterns of shared constructions in interaction and reveal that features such as their frequency and the amount of different constructions used for a referent are associated with the degree of object labelling convergence the participants exhibit after social interaction. More generally, the present study shows that automatically detected shared constructions offer a useful level of analysis to investigate the dynamics of reference negotiation in dialogue.

Abstract

In face-to-face dialogues, the form-meaning relationship of co-speech gestures varies depending on contextual factors such as what the gestures refer to and the individual characteristics of speakers. These factors make co-speech gesture representation learning challenging. How can we learn meaningful gestures representations considering gestures’ variability and relationship with speech? This paper tackles this challenge by employing self-supervised contrastive learning techniques to learn gesture representations from skeletal and speech information. We propose an approach that includes both unimodal and multimodal pre-training to ground gesture representations in co-occurring speech. For training, we utilize a face-to-face dialogue dataset rich with representational iconic gestures. We conduct thorough intrinsic evaluations of the learned representations through comparison with human-annotated pairwise gesture similarity. Moreover, we perform a diagnostic probing analysis to assess the possibility of recovering interpretable gesture features from the learned representations. Our results show a significant positive correlation with human-annotated gesture similarity and reveal that the similarity between the learned representations is consistent with well-motivated patterns related to the dynamics of dialogue interaction. Moreover, our findings demonstrate that several features concerning the form of gestures can be recovered from the latent representations. Overall, this study shows that multimodal contrastive learning is a promising approach for learning gesture representations, which opens the door to using such representations in larger-scale gesture analysis studies.

Abstract

Gestures are integral components of face-to-face communication. They unfold over time, often following predictable movement phases of preparation, stroke, and re-
traction. Yet, the prevalent approach to automatic gesture detection treats the problem as binary classification, classifying a segment as either containing a gesture or not, thus failing to capture its inherently sequential and contextual nature. To address this, we introduce a novel framework that reframes the task as a multi-phase sequence labeling problem rather than binary classification. Our model processes sequences of skeletal movements over time windows, uses Transformer encoders to learn contextual embeddings, and leverages Conditional Random Fields to perform sequence labeling. We evaluate our proposal on a large dataset of diverse co-speech gestures in task-oriented face-to-face dialogues. The results consistently demonstrate that our method significantly outperforms strong baseline models in detecting gesture strokes. Furthermore, applying Transformer encoders to learn contextual embeddings from movement sequences substantially improves gesture unit detection. These results highlight our framework’s capacity to capture the fine-grained dynamics of co-speech gesture phases, paving the way for more nuanced and accurate gesture detection and analysis.

Abstract

Sound and movement are entangled in animal communication. This is obviously true in the case of sound-constituting vibratory movements of biological structures which generate acoustic waves. A little less obvious is that other moving structures produce the energy required to sustain these vibrations. In many species, the respiratory system moves to generate the expiratory flow which powers the sound-constituting movements (sound-powering movements). The sound may acquire additional structure via upper tract movements, such as articulatory movements or head raising (sound-filtering movements). Some movements are not necessary for sound production, but when produced, impinge on the sound-producing process due to weak biomechanical coupling with body parts (e.g., respiratory system) that are necessary for sound production (sound-impinging movements). Animals also produce sounds contingent with movement, requiring neuro-physiological control regimes allowing to flexibly couple movements to a produced sound, or coupling movements to a perceived external sound (sound-contingent movement). Here, we compare and classify the variety of ways sound and movements are coupled in animal communication; our proposed framework should help structure previous and future studies on this topic.

Abstract

We present a dataset of behavioural and fMRI observations acquired in the context of humans involved in multimodal referential communication. The dataset contains audio/video and motion-tracking recordings of face-to-face, task-based communicative interactions in Dutch, as well as behavioural and neural correlates of participants’ representations of dialogue referents. Seventy-one pairs of unacquainted participants performed two interleaved interactional tasks in which they described and located 16 novel geometrical objects (i.e., Fribbles) yielding spontaneous interactions of about one hour. We share high-quality video (from three cameras), audio (from head-mounted microphones), and motion-tracking (Kinect) data, as well as speech transcripts of the interactions. Before and after engaging in the face-to-face communicative interactions, participants’ individual representations of the 16 Fribbles were estimated. Behaviourally, participants provided a written description (one to three words) for each Fribble and positioned them along 29 independent conceptual dimensions (e.g., rounded, human, audible). Neurally, fMRI signal evoked by each Fribble was measured during a one-back working-memory task. To enable functional hyperalignment across participants, the dataset also includes fMRI measurements obtained during visual presentation of eight animated movies (35 minutes total). We present analyses for the various types of data demonstrating their quality and consistency with earlier research. Besides high-resolution multimodal interactional data, this dataset includes different correlates of communicative referents, obtained before and after face-to-face dialogue, allowing for novel investigations into the relation between communicative behaviours and the representational space shared by communicators. This unique combination of data can be used for research in neuroscience, psychology, linguistics, and beyond.

Abstract

In this increasingly data-rich world, visual recordings of human behavior are often unable to be shared due to concerns about privacy. Consequently, data sharing in fields such as behavioral science, multimodal communication, and human movement research is often limited. In addition, in legal and other non-scientific contexts, privacy-related concerns may preclude the sharing of video recordings and thus remove the rich multimodal context that humans recruit to communicate. Minimizing the risk of identity exposure while preserving critical behavioral information would maximize utility of public resources (e.g., research grants) and time invested in audio–visual​ research. Here we present an open-source computer vision tool that masks the identities of humans while maintaining rich information about communicative body movements. Furthermore, this masking tool can be easily applied to many videos, leveraging computational tools to augment the reproducibility and accessibility of behavioral research. The tool is designed for researchers and practitioners engaged in kinematic and affective research. Application areas include teaching/education, communication and human movement research, CCTV, and legal contexts.

Abstract

In many musical styles, vocalists manually gesture while they sing. Coupling between gesture kinematics and vocalization has been examined in speech contexts, but it is an open question how these couple in music making. We examine this in a corpus of South Indian, Karnatak vocal music that includes motion-capture data. Through peak magnitude analysis (linear mixed regression) and continuous time-series analyses (generalized additive modeling), we assessed whether vocal trajectories around peaks in vertical velocity, speed, or acceleration were coupling with changes in vocal acoustics (namely, F0 and amplitude). Kinematic coupling was stronger for F0 change versus amplitude, pointing to F0's musical significance. Acceleration was the most predictive for F0 change and had the most reliable magnitude coupling, showing a one-third power relation. That acceleration, rather than other kinematics, is maximally predictive for vocalization is interesting because acceleration entails force transfers onto the body. As a theoretical contribution, we argue that gesturing in musical contexts should be understood in relation to the physical connections between gesturing and vocal production that are brought into harmony with the vocalists’ (enculturated) performance goals. Gesture–vocal coupling should, therefore, be viewed as a neuro–bodily distributed aesthetic entanglement.

Abstract

Conversational turn taking in humans involves incredibly rapid responding. The timing mechanisms underpinning such responses have been heavily debated, including questions such as who is doing the timing. Similar to findings on rhythmic tapping to a metronome, we show that floor transfer offsets (FTOs) in telephone conversations are serially dependent, such that FTOs are lag-1 negatively autocorrelated. Finding this serial dependence on a turn-by-turn basis (lag-1) rather than on the basis of two or more turns, suggests a counter-adjustment mechanism operating at the level of the dyad in FTOs during telephone conversations, rather than a more individualistic self-adjustment within speakers. This finding, if replicated, has major implications for models describing turn taking, and confirms the joint, dyadic nature of human conversational dynamics. Future research is needed to see how pervasive serial dependencies in FTOs are, such as for example in richer communicative face-to-face contexts where visual signals affect conversational timing.

Abstract

Do communicative actions such as gestures fundamentally differ in their control mechanisms from other actions? Evidence for such fundamental differences comes from a classic gesture-speech coordination experiment performed with a person (IW) with deafferentation (McNeill, 2005). Although IW has lost both his primary source of information about body position (i.e., proprioception) and discriminative touch from the neck down, his gesture-speech coordination has been reported to be largely unaffected, even if his vision is blocked. This is surprising because, without vision, his object-directed actions almost completely break down. We examine the hypothesis that IW’s gesture-speech coordination is supported by the biomechanical effects of gesturing on head posture and speech. We find that when vision is blocked, there are micro-scale increases in gesture-speech timing variability, consistent with IW’s reported experience that gesturing is difficult without vision. Supporting the hypothesis that IW exploits biomechanical consequences of the act of gesturing, we find that: (1) gestures with larger physical impulses co-occur with greater head movement, (2) gesture-speech synchrony relates to larger gesture-concurrent head movements (i.e. for bimanual gestures), (3) when vision is blocked, gestures generate more physical impulse, and (4) moments of acoustic prominence couple more with peaks of physical impulse when vision is blocked. It can be concluded that IW’s gesturing ability is not based on a specialized language-based feedforward control as originally concluded from previous research, but is still dependent on a varied means of recurrent feedback from the body.

Abstract

Gestures during speaking are typically understood in a representational framework: they represent absent or distal states of affairs by means of pointing, resemblance, or symbolic replacement. However, humans also gesture along with the rhythm of speaking, which is amenable to a non-representational perspective. Such a perspective centers on the phenomenon of vocal-entangled gestures and builds on evidence showing that when an upper limb with a certain mass decelerates/accelerates sufficiently, it yields impulses on the body that cascade in various ways into the respiratory–vocal system. It entails a physical entanglement between body motions, respiration, and vocal activities. It is shown that vocal-entangled gestures are realized in infant vocal–motor babbling before any representational use of gesture develops. Similarly, an overview is given of vocal-entangled processes in non-human animals. They can frequently be found in rats, bats, birds, and a range of other species that developed even earlier in the phylogenetic tree. Thus, the origins of human gesture lie in biomechanics, emerging early in ontogeny and running deep in phylogeny.

Abstract

How does communicative efficiency shape language use? We approach this question by studying it at the level of the dyad, and in terms of multimodal utterances. We investigate whether and how people minimize their joint speech and gesture efforts in face-to-face interactions, using linguistic and kinematic analyses. We zoom in on other-initiated repair—a conversational microcosm where people coordinate their utterances to solve problems with perceiving or understanding. We find that efforts in the spoken and gestural modalities are wielded in parallel across repair turns of different types, and that people repair conversational problems in the most cost-efficient way possible, minimizing the joint multimodal effort for the dyad as a whole. These results are in line with the principle of least collaborative effort in speech and with the reduction of joint costs in non-linguistic joint actions. The results extend our understanding of those coefficiency principles by revealing that they pertain to multimodal utterance design.