Rhythmic timing underlies a broad set of human behaviours, including music and dance. Critically, the emerging ability to produce an internally generated rhythm (spontaneous motor tempo; SMT), and adapt one’s movement to match an external stimulus (sensorimotor synchronisation; SMS), is related to success in language acquisition. Poor SMS is related to language difficulties in typically developing pre-schoolers (Carr et al., 2014; Politimou et al., 2019; Rios-Lopez et al., 2019), and across language disorders, including dyslexia (Thomson and Goswami, 2008; Lee et al., 2015; Persici et al., 2019), developmental language disorder (DLD; Corriveau and Goswami, 2009; Cumming et al., 2015), and speech impediments (Olander et al., 2010; Falk et al., 2015). Sensorimotor synchronisation is suggested as a useful tool for identifying those who may struggle with language, which is suitable from infancy and through the lifespan (Ladányi et al., 2020).

Since Fraisse (1982), laboratory measures of SMT in adulthood are often measured via tapping paradigms, where discreet intervals are produced by the vertical displacement of the index finger onto a keypad or equivalent surface. Whilst in adulthood the self-produced rate of tapping is stable within-subjects over short periods of time (Vanneste et al., 2001), across the lifespan SMT is known to change, with children’s tapping significantly faster than adult’s (McAuley et al., 2006). Adult SMT is in the range of 630 ms inter-onset-interval (IOI; McAuley et al., 2006). In very early childhood, SMT measured via tapping has been demonstrated as fast as 400–450 ms (IOI; Bobin-Bègue and Provasi, 2008). Whilst tapping dominates the SMT literature, and adult SMS studies, across early childhood, this difficult fine-motor task is not always the most appropriate. Particularly in determining infant ability to synchronize with external stimuli, different research groups have used a variety of tasks to facilitate synchrony within populations with poor fine motor skills.

Seminal studies into infant movement to music have simply allowed infants to move freely to auditory stimuli (Zentner and Eerola, 2010; Fujii et al., 2014). However, such scenarios do not provide auditory/haptic feedback equivalent to the tapping measures used in adulthood. Other infant paradigms used small hand-held instruments such as bells (Rocha and Mareschal, 2017). The closest experimental paradigm to tapping involves whole-hand drumming. From 5 months-of-age, infants can produce their own SMT via drumming (Rocha et al., 2021b) with the tempo and regularity of their drumming increasing over the first 2 years of life. Whilst infants cannot reliably synchronise their movements to music, a longitudinal investigation of infant drumming to nursery rhymes of different tempi suggests that by 11-months-of-age infants are beginning to shift away from their SMT to better match the rate of the song (Rocha et al., 2021a). Studies of toddlers evidence good tempo adaptation in older infants, when drumming along with a human and non-human partner (Kirschner and Tomasello, 2009; Yu and Myowa, 2021). In contrast to whole-body free movement analysis, constraints imposed in a drumming task allow more direct comparison of SMS over age, with a common effector and motion as is commonly used in adult tapping studies. As drumming can be used across contexts, with minimal apparatus and instruction, and from 5-months-of-age with no upper limit, we suggest that this could be a candidate marker of SMS that could be used at scale to detect early individual differences. Whilst there is a strong movement toward identification of risk of language disorders using neural markers, e.g., (Attaheri et al., 2022) an accessible behavioural assessment of rhythmic skill would have multiple practical advantages in identifying children at risk.

In the current study we test the feasibility of measuring SMS in the child’s own home, using asynchronous data collection methods, and largely automated data processing. If viable, such a technique will allow for large scale data collection. Our approach is focussed on creating an open source tool to evaluate rhythm in developing populations using a task that is low cost, easy to administer, and easy to adapt for research and clinical needs. Here we describe the implementation of our paradigm on a group of 2-year-olds, a notoriously difficult age to test, and document the successes and failures of our approach. We first ask whether we can detect the rate of drumming from home-video footage. We then characterise toddler Spontaneous Motor Tempo, and ask whether toddlers of this age show signs of successful SMS at a group level.

A total of 68 infants completed the experiment. Of these, 18 demonstrated no visible drumming and are excluded from further analysis. This included 14 who were partially off-camera, and a further four who presented no drumming in any trial. The remaining 50 infants provided drumming data in 208 different trials averaging 1.12 SMT trials each and 3.04 drumming trials across the different conditions. For each trial, drumming frequency was recorded for subsequent analysis.

Here we demonstrate the feasibility of using online measurement of infant drumming as an index of infant rhythmic skill. Infants in our sample showed age-appropriate tempo-flexibility, drumming faster to faster tempi and slower to slower tempi. Infants showed evidence of adjusting their rate of drumming away from their intrinsic rate of movement, or Spontaneous Motor Tempo (SMT). However, infants were not close to adult levels of tempo-matching (in the range of tens of milliseconds), showing an average mismatch of greater than 100 ms. This level of tempo-matching is in line with prior observations of 18-month-olds (Rocha and Mareschal, 2017). Previously, 24-month-olds have been shown to synchronise their drumming, but only when interacting with a live, social partner (see Kirschner and Tomasello, 2009; Yu and Myowa, 2021). Whilst our stimuli involved a video recording of a human hand drumming, it was not an overtly “social” signal. Nonetheless, our results show that it is possible to gain a behavioural index of infant sensorimotor synchronisation using a low-cost and accessible open-source platform.

Methodologically this work has multiple strengths. Firstly, by leveraging the Lookit platform, we were able to collect high resolution behavioural data with good ecological validity from a notoriously challenging age group, with minimal experimenter oversight. Because data were collected at home, infants were in a highly familiar setting and caregivers could run the study at time of their choosing. Parents could even abandon an attempt and try at a later time. Secondly, despite the variability of testing circumstances, we were able to get good compliance with instructions and engagement with the task. For example, in several cases older siblings were present but data could be screened for interference and distraction. Finally, the data processing pipeline provided objective measures of movement and rhythmicity with a relatively small amount of manual coding.

Our study demonstrates that markerless motion capture data with infants can be collected in a home with no specialist equipment. The data quality was sufficient for us to extract measures of infant motor tempo with automated Fourier transforms. This is a promising proof of concept, particularly given that the OpenPose model was trained primarily with adult data (Cao et al., 2021). The best infants participants produced data comparable to adult pilot participants (see “Supplementary Material”). It is important to observe that for infants with little or no drumming the Fourier method will not automatically extract a drumming frequency due to lower frequency noise. If additional manual coding was used to tag periods of drumming then more accurate tempo scores could be extracted. Future work will develop this functionality. Further, the trial lengths were purposefully short (20-s), in order to minimise attrition from the study in this unique testing scenario where the experimenter is not present, but this may not have allowed enough time for all infants to provide data. The toddlers mostly tolerated the length very well, and increasing the trial length to 1-min may allow more time for the infants to “warm up” to the drumming and produce enough data for analysis.

One general limitation of this approach to motion capture is that data is only two-dimensional, in the plane of the camera (X and Y coordinates). OpenPose does have the capability to combine data from multiple cameras to reconstruct three-dimensional poses and movement (Nakano et al., 2020). However, this only works in highly optimised conditions with a single participant in view, making it unsuitable for our web-based, parent supervised paradigm. However, the design of data processing workflow allows us to substitute alternative, improved pose estimation algorithms. See Wang et al. (2021) for a review of recent advances.

Future iterations of this paradigm can benefit from clearer instructions to both the caregiver and infant. Of the 68 infants recruited for the study, 18 did not provide drumming data in any of the experimental conditions. Further, the SMT exhibited in the first trial (SMT1) was slower than we would have predicted, and indeed slower than the SMT produced in SMT2, which was recorded at the end of the experiment. The caregivers were asked to prompt the infant “Can you drum for me?”, “Can you show me how you drum?”, but it is possible this was not sufficient for some infants to understand the task. Some parents reported that their child was unfamiliar with the word “drum,” while other infants simply responded “No!”. The original design did not include a video example, in an effort to not bias the infants’ SMT. However, a possible solution would be to include a video montage where two or more infants are drumming side-by-side, giving a clear demonstration of the action expected (repeated whole hand hits), whilst not giving a strong timing signal, adding clarity and motivation to participate. Further, if infants do not produce drumming during SMT1, it could be possible for the parent to replay the demonstration video, and the child attempt the trial again. This would better enable the collection of a representative SMT.

The largest source of missing data was poor camera angles that meant the infant hand could not be tracked (N = 14). Lookit is optimised for desktop/laptop computers (i.e., not tablets or mobile phones), and anecdotally, the angle of webcams is normally optimised for centring the adult face. The infant hand is considerably lower in the camera’s field. Prior to commencing the study, the caregiver was shown a preview of their camera angle and asked to check that their infant was in shot. However, as the infant was not yet drumming at this point, it may not have been obvious as to whether the area that would be drummed upon was in view. Future iterations can preview the view of the webcam between each experimental trial, such that the caregiver can adapt their angle as needed. However, care must be taken not to make the task instruction too complex or demanding for caregivers to follow.

A further limitation of remote asynchronous testing that we did not foresee was that not all children were sat at a table that provided good auditory and haptic feedback from the children’s drumming. Subtle differences in the surface infants were drumming on could potentially affect their ability to adapt their behaviour to the auditory stimuli we asked them to synchronise with. Future iterations could therefore also ask the parent and child to “sound check” their drumming to ensure they are getting auditory feedback when they hit the table. Further, if infants are recruited for a study outside of Lookit (e.g., where they are already participating in a lab based session), the paradigm could also be used either within the lab, or at home with a standardised surface (e.g., drum), provided by the research team, to ameliorate these differences.

The study provides promising basis for further exploration of other domains. Firstly, rhythmic movements of the whole body could be examined, allowing investigation of dancing and entrainment to music. OpenPose has been used in this context in laboratory studies with adults (Zeng and Chen, 2021). Our work demonstrates that this could be feasibly done with relatively large samples of young children. Even more promising would be to investigate synchrony between individuals, especially in the context of bonding and responsive caregiving where current human coding measures are labour intensive and lack standardization and predictive validity (Lotzin et al., 2015). Automated solutions have been a goal of social signal processing for a long time (Chetouani et al., 2017) and movement data has emerged as a promising signal (Egmose et al., 2017; López Pérez et al., 2017). However, progress has been slow and most methods are not suitable for field data (Chu et al., 2015). Therefore, we are currently adapting our methods for use in this context.

In summary, here we provide a successful proof-of-concept that we can extract the rate and accuracy of infant drumming from home video, using largely automated and fully open-source procedures. In an initial study of 2-year-old toddlers, we find evidence for tempo-flexibility, but not synchronisation, in response to an isochronous external beat presented at different tempi. The overarching goal of the current work was to develop a tool that is suitable to assess rhythmic movement in very young children, which can be employed at scale, and potentially even identify children at risk of neurodevelopmental disorders, including speech and language difficulties. Such longitudinal assessments, that are appropriate over developmental time, are key to understanding the mechanistic profiles of rhythm impairments across a broad range of neurodevelopmental disorders (Lense et al., 2021). Now that the feasibility of this online approach has been demonstrated, future work can refine the procedure, and further develop this promising tool for deeper insights into infant behaviour.

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by Goldsmiths, University of London. Verbal informed consent to participate in this study was provided by the participants’ legal guardian via a recorded video statement. Written informed consent was obtained from the minor(s)’ legal guardian, for the publication of any potentially identifiable images or data included in this article.

SR and CA co-designed the study, collected and analysed the data, wrote the manuscript, and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.