Developmental trajectories of vocal behaviors in common marmosets as a reference framework for neurobehavioral studies

Nonhuman primates are essential model animals for understanding brain development and neurodevelopmental disorders. The common marmoset is becoming a key primate species due to its ease of handling and reproductive efficiency, along with increasing research resources, such as an MRI brain atlas of developing marmosets. The present study aims to further support the use of marmosets by providing a detailed description of the developmental trajectories of vocal behaviors with an automated system to detect and analyze vocalizations. We collected 64,736 vocalizations, classified them into 11 vocal types, and observed significant changes in their vocal behaviors from neonates to 12 weeks of age. Several call types, including Trill, Phee, and Cry, peaked around the sixth week before gradually declining. By contrast, calls such as immature Phee, and Twitter were most frequent at birth but decreased steadily, nearly disappearing by 6–8 weeks. In addition, we provide a platform for automatically detecting these vocalizations using artificial neural networks trained on a dataset applicable across various research contexts. Overall, this study not only describes the developmental trajectories of vocal behaviors in a fixed controlled context across developmental stages, but also serves as a reference framework for analyzing marmoset models of developmental disorders.

Nonhuman primates (NHPs) constitute essential animal models in neuroscience because of their close phylogenetic relationship to humans, and among them, the common marmoset (Callithrix jacchus) has become an important species (Feng et al., 2020; Miller et al., 2016; Mitchell and Leopold, 2015; Samandra et al., 2022). Marmosets offer several practical advantages, such as easy handling, relatively rapid development, and efficient reproduction (Kishi et al., 2014), which allow them to complement other NHPs like macaques. These advantages have enabled marmosets to serve as a valuable platform for adapting modern molecular technologies, primarily developed in rodents, for use in NHPs. Because such optimization typically requires several iterations and a relatively large number of animals, marmosets provide an efficient bridge across species. Successful advances in marmosets, including calcium imaging (Kondo et al., 2018; Sadakane et al., 2015), the development of cell-type-specific promoters (Matsuzaki et al., 2024; Shinohara et al., 2016), and the production of gene-modified animals (Park et al., 2016; Sasaki et al., 2009; Sato et al., 2016), have in turn facilitated the broader application of these techniques to macaques.

Furthermore, marmosets are a valuable model for studying developmental processes of social and communicative behaviors and their disruption (Phaniraj et al., 2026). Marmosets have a variety of call repertoires (Agamaite et al., 2015) and exhibit complex communication dynamics through their calls (Grijseels et al., 2024; Oren et al., 2024). Increasing research has revealed vocal plasticity during development, in which vocal behaviors are jointly shaped by interactions between internal maturational processes and possible external factors (Biazzi et al., 2025; Gultekin and Hage, 2017, 2018; Gultekin et al., 2021; Qi et al., 2025; Takahashi et al., 2016, 2017). For example, neonate marmosets produce distinctive immature calls, and social feedback from their parents shapes the trajectory of vocal maturation (Hage, 2020). Research on vocal development in marmosets may offer valuable insights into human vocal learning and the emergence of language. In addition, it will become increasingly important to evaluate the developmental trajectories of animal models for developmental disorders. A seminal study has described altered vocal features, along with changes in cortical synaptogenesis, synaptic function, and gene expression in a marmoset model of autistic spectrum disorder induced by valproic acid treatment (Watanabe et al., 2021).

However, there is still a lack of behavioral paradigms and analytical frameworks to evaluate developmental processes in model marmosets. There are growing research resources for marmosets, including databases of gene expression (Kita et al., 2021; Shimogori et al., 2018) and neuronal connectivity (Majka et al., 2020a; Watakabe et al., 2023), as well as a brain atlas of MRI and histology (Hata et al., 2023; Liu et al., 2020; Majka et al., 2020b; Woodward et al., 2018). By contrast, only a few technical platforms for the behavioral domain are available (Kaneko et al., 2024). The detection and classification of vocalizations in NHPs relies primarily on manual annotation (Yano-Nashimoto et al., 2024), and automated methods using artificial neural networks have recently emerged as a viable option (Lamothe et al., 2025; Oikarinen et al., 2019; Turesson et al., 2016; Wu et al., 2024; Zhang et al., 2018). Although such a computerized system would be particularly beneficial for studying marmoset development, as it requires longitudinal observations and analysis of a substantial number of calls, its application remains relatively scarce.

To address this issue, we aimed at establishing an automated system to analyze the vocal behaviors of developing marmosets. In the present study, we observed processes of vocal development over the first three postnatal months in 12 marmosets and manually annotated a total of 13,270 calls to train neural networks that can accurately detect a variety of marmoset calls. Our results provide a reference framework for studying the vocal development in marmosets and, also, for assessing that of their disease models.

Animals

Vocal development was recorded in 12 neonatal marmosets (6 males and 6 females) derived from 9 families. All animal care and experimental procedures were approved by the Animal Welfare and Animal Care Committee of the Center for the Evolutionary Origins of Human Behavior at Kyoto University, and were conducted in accordance with the Guidelines for the Care and Use of Nonhuman Primates established by the same institution. Infant marmosets were housed with their parents, without any food or water deprivation. The subject marmosets in the present study served as a control group in a separate study, providing a reference for genetically modified disease model marmosets. The subject animals were injected intravenously through the tail vein on the day of birth or the next day with the variant of AAV9 vector, AAV.CAP-B10 (Goertsen et al., 2022), which incorporates a scrambled miRNA sequence under the control of the CAG promoter.

Vocal recording experiment

The recording booth was a 90-cm cubic box with transparent acrylic walls and a metal mesh floor and ceiling (Figure 1A). The booth was divided into two spaces by a vertically oriented transparent acrylic baffle. Visual, auditory and olfactory contacts were available between animals. Two microphones were placed above the two compartments outside the booth. The male parent was kept in a small metal-mesh cage on the left side of the recording booth. The neonates were gently separated from the parents, kept in a dark area for 1 min, and then placed at the center of the right compartment of the recording booth. The recording duration was 6.5 min. Audio files were recorded at a sampling rate of 44.1 kHz. We recorded vocalizations three sessions per week during the first month of the birth, at least twice per week during the second month, and once per week thereafter up to about 4 months old.

Automated detection system

The neural networks used for vocal analysis consisted of two distinct networks (Figure 1B). The first network detected the timing of calls, irrespective of call types, across all recording sessions, and another network classified call types for each detection. For both networks, we used the analysis system named “DeepSqueak” (Coffey et al., 2019), which was initially developed for the analysis of rodents’ vocalizations and was recently used for the study of primate vocalizations (Romero-Mujalli et al., 2021). The two networks were trained via supervised learning using manually annotated datasets.

Infant and adult calls were discriminated based on the relative call amplitudes recorded by two microphones. A call was attributed to either the infant or the parents when the amplitude difference between the two microphones exceeded a threshold (Figure 1C). Around 5.2% of the calls were identified as the parents’ calls and excluded from the following analysis (Figure 1D).

Datasets for training and quality check of automated detection

A total of 13,270 vocalizations were manually labeled across 45 pseudo-randomly selected recording sessions to comprehensively cover a variety of vocalizations. To construct the training datasets for automated detection, in addition to the 12 marmosets described before, we have additional data from 14 marmosets with gene knockdown, in which a scrambled miRNA sequence in the virus vector was replaced by those for the knockdown of the target gene, such as POGZ (Assia Batzir et al., 2020; Matsumura et al., 2020; Suliman-Lavie et al., 2020) or CHD8 (Katayama et al., 2016; Kawamura et al., 2021). The calls were annotated with timing and frequency ranges on the spectrogram and classified into 11 types (Trill, Phee, Twitter, Cry, Immature Phee (Im-Phee), Trill-Phee, Trill-Twitter, Trill-Cry, Tsik, See, Ek) (Figure 1E).

Statistical test

A two-way ANOVA was performed to evaluate the developmental changes across ages and the differences in call types. The statistical model included the age and call type as main effects, the interaction between these, and subjects as a random factor. The α was set at 0.05.

Manual annotations for network training

We collected 231 recording sessions over the postnatal 3 months in 12 marmosets to evaluate vocal development. For neural network training, we have annotated 45 sessions (including additional gene-knockdown animals), resulting in 13,270 calls. Among them, 40 sessions with 11,405 calls were used for network training, and the rest were used for performance evaluation of the trained networks (Table 1). The recording sessions used for manual annotations covered most call types; however, two call types (See and Ek) were excluded from later analysis because of small sample sizes.

Accuracy of the detection network

Manually detected 11,405 calls (Table 1, Training data sets) were utilized to train the detection network. After 24,100 iterations (100 epochs) of training, the detection network achieved a precision of 0.923, a recall of 0.887, and an F1 score of 0.905 A true positive was defined as having an intersection over union (IoU) ≥ 0.5.

Accuracy of the classification network

Manually annotated 11,405 calls (Table 1, Training data sets) were used to train the classification network up to 89,000 iterations (1,000 epochs). During training, 90% of the data were used to update the network weights, and the remaining 10% were used as a validation dataset to determine the optimal number of training iterations. Then, the network state at 5,020 iterations (epoch 57) achieved the highest validation accuracy of 89.1% (Figures 2A,B).

Using a quality-check dataset that is independent of the training dataset, we estimated F1 scores for nine call categories, except for Ek and See, for which there are only a few manual annotation samples (Figure 2C). Some call types showed good classification accuracy, though considerable variation is evident across them. Notably, the confusion matrix (Figure 2B) showed that the classification of compound vocalizations was a significant source of error in the automated detections. For example, there were several erroneous classifications between Trill and Trill-Phee, and between Phee and Trill-Phee (similarly for Cry and Twitter). The F1 scores of all compound calls were lower, particularly for Trill-Cry, which had the fewest samples in the training dataset. We classified the calls into 11 discrete categories. However, there were several complex combinations of acoustic motifs, and there was continuity across calls in different classes. These facts might require an alternative approach to classification based on the definitions of human observers (also see section “4. Discussion”).

To provide a practical and straightforward solution to evaluate developmental trajectories along with the current approach, we combined compound calls with their corresponding individual calls into the same categories (Figure 2D). Then, we simplified the call types to six categories, and obtained F1 scores for each call type (Figure 2E). Overall F1 scores for all call types were over 0.8, ensuring objective and precise evaluation for longitudinal observations for developing marmosets.

To further assess the reliability across developmental stages, we randomly selected 7 sessions from the first 5 weeks (early stage) and the last 5 weeks (later stage), and obtained an F1 score for each group. The detection network extracted unclassified calls from these sessions. Then, these calls were manually classified into call types and compared with the output of the classification network. The results indicated that the classification performance was comparable among age groups, suggesting no systematic bias across developmental stages (Figures 2F,G).

Developmental trajectories as assessed by automated detection

We then analyzed 231 recording sessions from 12 marmosets to quantify vocal development over 3 months, and the networks identified 64,736 calls. With advancing age, the overall number of vocal behaviors decreased over 15 weeks (Figure 3A) [the main effect of age, F (13,715) = 5.31, p < 0.0001]. A distinctive pattern of developmental trajectories has been observed in different calls [the interaction between age and call type, F (65,715) = 4.63, p < 0.0001]. The contact calls, including Phee, Twitter, and Trill, showed similar unimodal distributions, although the time-to-peak varied across call types: Phee and Trill reached peak calling in the 6th–7th week; thereafter, Phee decreased, whereas Trill maintained its peak for another 5 weeks. Twitter reached its peak in the 2nd week, much earlier than the others. The alarm call (Tsik) monotonically decreased over time. Calls typical of neonates showed a different tendency: Im-Phee bottomed out in the 6th week (while Phee reached its peak); Cry exhibited a bimodal pattern, with peaks at weeks 2 and 7, ultimately approaching a negligible level (Figure 3B).

This study aims to establish an efficient and user-friendly method for automated analysis of the development in vocal behaviors of marmosets. We successfully detected marmoset vocalization patterns using DeepSqueak, originally designed for detecting rodent ultrasonic vocal, and all procedures were available via the GUI without modifying the original code. Most calls were detected with high accuracy, except compound calls, which were difficult to distinguish due to their small sample sizes and acoustic similarity. Therefore, we combined the compound calls with their corresponding individual calls (see Figure 2E), thereby improving the reliability of the detection results. Under these conditions, we could elucidate the developmental changes in six different call types. The calls detected by the networks were sufficiently precise, and overall developmental patterns were consistent with those reported in previous studies by manual annotations (Agamaite et al., 2015; Gultekin et al., 2021; Takahashi et al., 2015; Yano-Nashimoto et al., 2024).

Alterations in call distribution may reflect the developmental changes of two distinct but complementary aspects. The first change is the infant affective state during the recording sessions. Infant marmosets spend most of their time being carried by their parents until 4–6 weeks of age, and begin locomoting independently as they grow (Schultz-Darken et al., 2016). This transition is not entirely voluntary but is also shaped by parental behavior, as parents may refuse to carry the infant even when it attempts to climb onto their backs. This process may induce separation anxiety and evoke more calls. Consistent with this view, Cry call, typically used to attract parental attention or elicit parental approach (Ziegler et al., 2017), reached another peak at this age. Subsequently, infants may become less anxious as they get older. Such a developmental change in the affective state was also likely reflected in alarm calls (i.e., Tsik) of which acoustic parameters remained largely invariant across ages, and the frequency of these calls progressively diminished (see Figure 3B). The second change is the transition from immature to mature calls, as evidenced by the inverse trend in Im-Phee and Phee (see Figure 3B). This transition may reflect the physical development of organs involved in vocal production, such as lung and larynx (Takahashi et al., 2015).

It should be noted here that the infant vocalization pattern for some call types may change depending on the experimental protocol in laboratory environments. For example, the increase in Trill calls is prominent in our study during 6–12 weeks, while it is moderate (Yano-Nashimoto et al., 2024) or absent in other studies (Pistorio et al., 2006; Qi et al., 2025; Takahashi et al., 2015). These differences can partially be explained by visual contact between animals. Trill is a contact call used at relatively close distances between animals (Landman et al., 2020), whereas most studies about infant vocal recordings do not allow visual contact with the parents. Similarly, the second peak of Cry calls after 5 weeks was observed in some studies (Gultekin et al., 2021; Yano-Nashimoto et al., 2024) as well as in ours, whereas no such transient increases were observed in others (Qi et al., 2025; Takahashi et al., 2015). It is unclear what experimental parameters might cause these differences, as various studies employ slightly different protocols. Even using the same protocol, individual differences in Cry are high around this age, and, therefore, some parameters in some experimental protocols may exaggerate individual differences arising from the developmental plasticity (Gultekin et al., 2021; Yano-Nashimoto et al., 2024).

The developmental trajectories of neonatal marmoset calls obtained from the automated detection pipeline provide a reference for addressing the developmental process of disease model animals. Impaired language ability often occurs in patients with neurodevelopmental disorders. For example, in patients with White-Sutton syndrome, a recently identified genetic disorder with a wide spectrum of neurocognitive symptoms, speech delay is one of the most universal symptoms (Assia Batzir et al., 2020). The marmoset is an ideal translational model for studying the underlying brain mechanisms (Miller et al., 2016; Samandra et al., 2022; Schultz-Darken et al., 2016). The substantial number of annotations and high analytical precision achieved in the present study offer an efficient framework and a reliable reference for evaluating the vocal development in disease model marmosets.

There are three aspects that should be considered when using the present protocol to evaluate the developmental trajectories of model animals. First, as argued above, the developmental changes in call behavior might reflect multiple components, such as motor development or affective responses to the experimental condition. Thus, interpretations of the behavioral phenotype of model animals should require careful assessment of the mechanisms underlying the behavioral changes. Second, the networks obtained in the present study were optimized for infants or juveniles under 6 months of age, but not for adults. Last, while the present study classified calls into discrete categories, this approach is still insufficient to fully describe the development in marmoset vocal behaviors. Boundaries between calls in different categories may be ambiguous, as shown in Figure 2. On the other hand, the calls within an identical category have variations in their acoustic parameters (Agamaite et al., 2015; Pistorio et al., 2006). A desirable approach could evaluate such continuity over categories and variations within single categories. Recent progress in generative artificial intelligence might provide an alternative approach, such as a Variational Autoencoder (VAE), that is able to capture the complex nature of calls by using a few parameters, thereby enabling quantitative analysis of subtle changes across similar calls (Agamaite et al., 2015), and is used, for example, to assess an atypical vocal development in a rat model of schizophrenia (Potasiewicz et al., 2025). Application of such an approach to the NHP model is worth addressing in a future study.