Utterance evolution: the road to generative, combinatorial communicators Authors: Catherine Crockford, Roman Martin Wittig, Cedric Girard‐Buttoz Published: 2026 ABSTRACT ABSTRACTLanguage has long been considered uniquely complex in the animal kingdom; however, animal research over the last decade has begun to challenge some long‐standing premises about exactly which language capacities are uniquely human. The task of resolving why and how complex communication systems evolve, particularly human language, has manifested in a rich multi‐disciplinary field, which relies on accurate empirical and comparative assessment of natural animal communicative capacities. For example, comparative models of relevant gene and brain functions will remain limited without cracking the animal communication code. To achieve this, empirical research needs an updated, quantitative and comparative framework that distinguishes combinatorial systems from simpler ones. We offer such a framework by integrating new work, demonstrating diverse meaning‐bearing combinatorial signalling strategies in natural animal communication, into cross‐disciplinary theoretical frameworks. We detail four major mechanistic transitions which we argue offer both evolutionary and ontogenetic routes from non‐combinatorial communicators, that mainly rely on single signal utterances, to generative combinatorial signal users, that mainly rely on combinatorial, meaning‐based utterances. These transitions are measurable and predictable: (i) signal combination versatility; (ii) signal combination learning and conventionalization; (iii) message expansion through combinatorial mechanism diversity and versatility; and (iv) message expansion through syntax involving linear or hierarchical rules. We offer testable hypotheses for each transition and propose that species demonstrating combinatorial Transitions (i) and (ii) will be prime candidates for testing Transitions (iii) and (iv), where generativity can emerge. We also point to future methodological tools to assess the meaning of signals and their combinations in signal utterance‐to‐context mapping analyses. Critically, we argue that simply having research on different species is not enough – testing evolution of language theories requires comparative research across species and ontogeny, based on relatable, quantitative measures. To date, only humans demonstrate meaning‐based, generative communication. However, recent discoveries, particularly of combinatorial versatility in chimpanzee (Pan troglodytes) and bonobo (Pan paniscus) vocal sequences, open the possibility of generative communication in other species. Whether such traits are human‐unique can no longer be assumed but must be tested.Keywords: Vocal sequences, call combination, compositionality, syntax, meaning, language evolution, vocal learning, animal communication, bioacoustics, mammals, birds, chimpanzees INTRODUCTION Language has long been considered uniquely complex in the animal kingdom, largely due to its combinatorial and generative nature (Hauser, Chomsky & Fitch, 2002; Fitch, Huber & Bugnyar, 2010; Friederici, 2011; Suzuki, Wheatcroft & Griesser, 2018; Zuberbühler, 2019; Schlenker et al., 2023), where words can be endlessly combined and recombined to create an infinite range of new meanings. To claim that traits are uniquely human requires empirical assessment of other animal signalling systems. However, there is a dearth of empirical tests that assess whether other animals possess generative, combinatorial capacities, and with good reason. Decoding animal signal systems is a cumulative process. For example, we need to understand what single signals mean before we can understand the information conveyed by signal combinations. This problem is well illustrated by Hauser et al. (2002) who laid out numerous relevant comparative approaches to test which components of the faculty of language animals possess, but none of them explicitly examined combinatorics or generativity in natural animal systems. Recent advances in ethological research in natural animal signal systems allow us to dig into how, how often and how much animals combine signals that modify, add or change meaning (Suzuki, Wheatcroft & Griesser, 2016; Engesser, Ridley & Townsend, 2016; Townsend et al., 2018; Suzuki et al., 2018; Engesser & Townsend, 2019; Zuberbühler, 2019, p. 218; Girard‐Buttoz et al., 2022b; Schlenker et al., 2023; Berthet, Surbeck & Townsend, 2025; Girard‐Buttoz et al., 2025), revealing unexpected and widespread combinatorial capacities in signalling (Fig. 1). In spite of these advances, no non‐human animal system is considered inherently combinatorial or generative. This is partly due to a lack of integration of empirical and theoretical work, and partly due to difficulties comparing capacities across species. Here, we integrate both recent empirical (reviewed in Schlenker et al., 2016, 2024; Townsend et al., 2018; Berthet et al., 2022) and past theoretical work into a framework which facilitates quantitative comparison across phylogeny and ontogeny. The aim is to aid identification of combinatorial and generative animal signalling systems (Freeberg, Dunbar & Ord, 2012; Manser et al., 2014; Townsend et al., 2018; Peckre, Kappeler & Fichtel, 2019; Zuberbühler, 2019; Schlenker et al., 2023). Throughout this review, we focus on animal signal combinations, including vocal or non‐vocal modalities. Combinatorial mechanisms documented in non‐human animal natural signalling systems. Each mechanism has so far been reported only in call combinations comprising 2 units (bigrams). The first column states how meaning is modified in bigrams compared to the constituent single units. The second column indicates the linguistic terminology used in the animal communication literature. The third column presents a theoretical formulation of each mechanism. The fourth column uses silhouettes to depict the non‐human animal species for which each phenomenon has been described (from top to bottom and left to right: Putty‐nosed monkeys (Cercopithecus nictitans), Arnold & Zuberbühler, 2008; bonobos (Pan Paniscus), Berthet et al., 2025; chimpanzees (Pan troglodytes), Girard‐Buttoz et al., 2025; Campbell's monkey (Cercopithecus campbelli), Arnold & Zuberbühler, 2006; titi monkeys (Callicebus nigrifrons), Berthet et al., 2019; black‐capped chickadee (Poecile atricapillus), Templeton, Greene & Davis, 2005; Freeberg, 2008; chimpanzees, Girard‐Buttoz et al., 2025; olive colobus (Procolobus verus), Gallot et al., 2025; bonobos, Schamberg et al., 2016; Diana monkeys (Cercopithecus diana), Coye, Zuberbühler & Lemasson, 2016; Japanese tits (Parus minor), Suzuki, Wheatcroft & Griesser, 2016; chimpanzees, Leroux et al., 2021; Girard‐Buttoz et al., 2022a, 2025; pied babblers (Turcoides bicolor), Engesser et al., 2016; banded mongooses (Mungos mungo), Jansen, Cant & Manser, 2012; Japanese tits, Suzuki et al., 2016; pied babblers, Engesser et al., 2016 and chimpanzees, Girard‐Buttoz et al., 2025). The last column provides specific examples of each mechanism (from top to bottom: Putty‐nosed monkeys (Arnold & Zuberbühler 2008); chimpanzees (Girard‐Buttoz et al., 2025), Campbell's monkeys, Arnold & Zuberbühler, 2006; chimpanzees, Girard‐Buttoz et al., 2025; Diana monkeys, Coye, Zuberbühler & Lemasson, 2016; chimpanzees, Leroux et al., 2021; Girard‐Buttoz et al., 2025, and Japanese tits, Suzuki et al., 2016. Whilst studies using alarm contexts form the foundation of our understanding of combinatorial systems, recent studies also demonstrate rich combinatorial usage in other contexts, such as during social interactions. A first step to understanding signal combinations is identifying single signals and their usage. A single signal constitutes a functional communicative unit. This can be a single call, a repetition of the same call type or a sequence of different calls (see Salis, Martin & Girard‐Buttoz, 2025, for a recent discussion on this issue). We define a signal combination in the broad sense, as the combination of at least two different meaning‐bearing signals into a longer meaningful structure without presupposition as to the relationship between the meaning of the composing signals and that of the combination (Table 1). Note that this definition differs from some stricter definitions found in the literature where combinatoriality is only characterized if the meaning of the combination is not a simple addition of the meanings of the composing signals (e.g. Scott‐Phillips & Blythe, 2013). Our semantic choice aligns with recent discussion of animal combinatorial communication (Engesser & Townsend, 2019) and allows a more inclusive consideration of animal communicative systems across a variety of taxa (Fig. 2). Table 1Glossary of terms, operationally defined to enable comparative quantitative analyses.Non‐combinatorial communicatorMainly relies on single‐signal utterances.Combinatorial communicatorMainly relies on combined utterances that expand meanings such that signal combinations are commonly used across daily life.Signal sequenceA generic term for signal output containing more than one signal type.Signal combinationA signal sequence in which at least two different meaning‐bearing units are strung in a sequence to form a meaningful composite.Combinatorial mechanismThe function by which meaning is changed in a signal combination compared to its constituent parts, such as modifying or adding meanings, or generating new meaning (Fig. 1). Combinatorial mechanisms can give rise to specific rules, such as in English, ‘un‐’ operates as a negation when prefixed to a word like ‘sympathetic’, or in Campbell's monkeys ‘oo’ down‐grades urgency when suffixed to alarm calls. Such mechanisms and rules may or may not overlap with linguistic rules.GenerativeCombinatorial mechanisms or rules that generalize across call types and contexts, such that each can liberally generate new messages or relationships.Vocal production learningAdding a new call to one's repertoire through individual or social learning.Vocal usage learningLearning to use an existing call in the same context as others use it, in a new context, or in the same sequential order as others use it, through individual or social learning.Signal formThe shape (or acoustic structure) of a signal.Signal functionThe communicative purpose of the signal, often ascertained through natural contextual usage or experiments.ConventionalizationSocial learning standardizes signal combination meanings within social groups or populations.SyntaxLinear or hierarchically structured rules or mechanisms governing meaning‐bearing signal combinations.Non‐adjacent dependenciesDependencies between nonadjacent elements within a sequence. Glossary of terms, operationally defined to enable comparative quantitative analyses. A comparative view of combinatorial signalling capacities across phylogeny and ontogeny in natural animal systems. We define signal combinations as those that demonstrate a shift or addition of meanings compared to the constituent signals. Combinatorial mechanisms describe how meaning changes, such as through meaning addition or modification (see Table 1). For decades, non‐human primates' apparently limited vocal skills have proved a conundrum for theories of how language evolved. Recent evidence puts primates, particularly extant hominids, ‘back in the game’ as they currently show the strongest evidence for language precursors, including generalized capacities to encode meaning within signal combinations, signal usage learning within signal combinations, and vocal sequence expansion through ontogeny. Circle: evidence of skill in vocal modality; asterisk: evidence in gestural modality. Light to dark circles: implicated or demonstrated evidence for limited (light), intermediate (bright) or expanded (dark) capacities. From the bottom and clockwise the species depicted are as follow: Pied babbler (Turcoides bicolor), black‐capped chickadee (Poecile atricapillus), Japanese tit (Parus minor), African elephant (Loxodonta africana), bottlenose dolphin (Tursiops truncatus), dingo (Canis lupus dingo), meerkat (Suricatta suricatta), banded mongoose (Mungos mungo), marmoset (Callithrix jacchus), titi monkeys (Callicebus nigrifrons), olive colobus (Procolobus verus), Campbell's monkey (Cercopithecus campbelli), Diana monkey (Cercopithecus diana), Putty‐nosed monkey (Cercopithecus nictitans), human (Homo sapiens), chimpanzee (Pan troglodytes) and, bonobo (Pan paniscus). References for the specific studies demonstrating these capacities in each species can be found in the main text and in Fig. 1. What makes language complex is layered flexibility in combining sounds, termed duality of patterning, where meaningless sounds used in speech (phonemes) are combined into words that bear meaning, which are themselves combined into hierarchically organized meaningful utterances (Hockett, 1959). Utterances flexibly describe relations between an endless array of real‐world objects and events, including complex social scenarios (e.g. Stewart, McElwee & Ming, 2013). Most languages use a small set of less than 50 phonemes (Moran, McCloy & Wright, 2012), which overlaps with the number of sounds used in other primate and animal vocal repertoires (see Section II). Hence, we have two expectations for any non‐human animal generative signalling system. First, it will rely on combinatorial capacities rather than large sound sets. Second, learning of signal‐to‐meaning mappings may be evident, and so benefit from ontogenetic assessment. The vast majority of animal utterances are composed of single calls, such as tweet, chirp, hoo, grunt, or repetitions of the same single call (e.g. Bouchet, Blois‐Heulin & Lemasson, 2013), some of which convey meaning, beyond the signaller's emotional state, to conspecifics about objects or events external to the signaller (see Section II). When such meaning‐bearing calls are combined into call combinations, animal examples to date (Fig. 1) demonstrate that meanings can be (i) modified such that one call modifies the meaning of the other(s) (Coye, Zuberbühler & Lemasson, 2016; Girard‐Buttoz et al., 2025), (ii) added such that the meaning of the composing calls is preserved (Suzuki et al., 2016; Engesser et al., 2016; Dutour, Lengagne & Léna, 2019; Girard‐Buttoz et al., 2025), or (iii) changed such that the combination generates an entirely new meaning unrelated to the meaning of the composing calls (Arnold & Zuberbühler, 2008; Girard‐Buttoz et al., 2025). We define a combinatorial mechanism as the function by which the meaning of a single signal is altered when strung into a signal combination, that is (i), (ii) and (iii) above are three distinct combinatorial mechanisms (Table 1). Such combinatorial signalling is evident both in species that can learn vocalizations (song birds: tits, Suzuki et al., 2016; Dutour et al., 2019; magpies, Walsh et al., 2019; elephants, Hedwig & Kohlberg, 2024) and in species considered to have more fixed, inherited vocal repertoires [primates: monkeys, Arnold & Zuberbühler, 2006, 2008; Coye et al., 2015, 2016; Berthet et al., 2019; Gallot et al., 2025; great apes, Schamberg et al., 2016; Schamberg, Surbeck & Townsend, 2024; Leroux et al., 2023; Berthet et al., 2025; Girard‐Buttoz et al., 2025; carnivores: meerkats (Suricata suricatta), Rauber, Kranstauber & Manser, 2020; banded mongooses (Mungos mungo), Jansen, Cant & Manser, 2012; babblers, Engesser et al., 2016; Figs 1 and 2]. Hence, it is now clear that combinatoriality is a general animal phenomenon used to expand the range of information that can be conveyed beyond the number of available single signals. However, for most species, signal combinations used are few and short, containing two call types (bigrams). With the exception of chimpanzees (Pan troglodytes) (Girard‐Buttoz et al., 2025), to date, no more than two combinatorial rules or mechanisms (Table 1; Fig. 1) have been documented per species, and these examples are mostly found in alarm contexts. Hence in most animal systems, whilst combinatoriality is used it does not liberally expand messaging. Critically, rigorous demonstrations of the capacity to combine meanings into utterances to describe relations between two or more concomitant real‐world objects and events in animal communication are few (Suzuki et al., 2016; Engesser et al., 2016; Girard‐Buttoz et al., 2022a; Bortolato et al., 2023a; Leroux et al., 2023; Girard‐Buttoz et al., 2025). Until utterances combine meanings, signallers remain highly limited in their capacity to signal about relations between agents, patients and events. Per species, the extent to which combinatorial capacities are flexibly utilized across daily life determines generative capacity – and the complexity – of the communication system. Species that are non‐combinatorial communicators have no or only few, short signal combinations utilizing only one combinatorial mechanism, and therefore have no generative capacity. In contrast, species which are combinatorial communicators would have many, signal combinations of various lengths used liberally across daily life events, and which can utilize several combinatorial mechanisms across call types and contexts (Table 1). Such a system would also have generative potential. Recent research suggests that, to some extent, the combinatorial limitations described per species reflect the limits of research effort rather than limitations in animal signalling combinatoriality (Berthet et al., 2025; Girard‐Buttoz et al., 2025). Major evolutionary theories tackle what drivers might promote the evolution of complex communication, such as the social complexity hypothesis for communicative complexity. This hypothesis posits that increasingly complex social life selected for signal complexity to provide more diverse and nuanced messages (Jolly, 1966; McComb & Semple, 2005; Freeberg et al., 2012; Manser et al., 2014; Peckre et al., 2019; Planer & Sterelny, 2021; Grampp et al., 2023). This is the case whether or not complexity approaches a human language‐like system. Research examining evolutionary drivers of signalling is hampered by a lack of clarity about what broadly defines or constitutes more versus less complex communication (Freeberg et al., 2012; Manser et al., 2014; Peckre et al., 2019), and complex systems more generally (Kappeler, 2019). We argue that a central tenet of signal complexity should be to expand messaging. In this review, however, we focus not on why, but on how, communicative complexity arose. The only known natural system to generate expansive messaging is language. Language evolved as a generative combinatorial system constructed from a small sound set used in speech – and less commonly utilizing a larger signal set in the case of sign language. (Hauser et al., 2002; Jackendoff & Wittenberg, 2017; Table 1). Whilst theories of language evolution are numerous, there is strong agreement that critical evolutionary transitions towards language involved acquiring capacities in meaning‐based combinatoriality, signal combination learning, and in the production and comprehension of syntactical, hierarchical structures (Lewis, 1969a; Nowak, Plotkin & Jansen, 2000; Hauser et al., 2002; Fitch et al., 2010; Skyrms, 2010; Petkov & Jarvis, 2012; Scott‐Phillips & Blythe, 2013; Nowicki & Searcy, 2014; Truswell, 2017; Jackendoff & Wittenberg, 2017; Townsend et al., 2018; Engesser & Townsend, 2019; Zuberbühler, 2019; Jarvis, 2019; Planer & Sterelny, 2021). How these components interconnected through evolution remains highly debated, mainly due to a dearth of empirical data to draw on. Insights from recent empirical discoveries in animal combinatorics can re‐focus the theory. Hence, this review has two goals. First, we propose a framework that specifically focuses on four major transitions that expand meaning‐based signal combinatoriality and signal combination learning. We argue these transitions offer both evolutionary and ontogenetic routes from mainly rudimentary, single‐signal communicators to versatile, combinatorial, generative communicators. Second, we identify key capacities on which to focus future research efforts to identify combinatorial and generative communicators. We point to measures that may facilitate empirical comparative and ontogenetic testing of how combinatorial and generative a given communicative system is. We offer a specific focus on the latter, given that, to date, only humans demonstrate meaning‐based signal generativity. In non‐vocal modalities, such as gesture and facial expression, or multimodal signalling, examples of meaning‐bearing sequences where combinations change or add meanings are currently scarce (Hobaiter & Byrne, 2011; Graham, Furuichi & Byrne, 2020; Amici, Oña & Liebal, 2024). Hence, we focus on the vocal domain, however, the principles discussed apply equally to sequences of any signal modality. To demonstrate this, we use the term ‘signal’ where suitable, but examples mainly detail vocal signals. Previous research (Nowak & Krakauer, 1999; Nowak et al., 2000; Nowak & Komarova, 2001) has proposed mathematical models of language evolution that assessed the conditions under which combinatorial signalling systems are selected over the use of single signals. These models demonstrated that a generalized combinatorial signalling system is likely to emerge when (i) the number of messages to be conveyed considerably surpasses the number of single signals in the repertoire; and (ii) the same signal can be reused in numerous combinations. Whilst (i) details the selection pressure, the need to communicate about a wider range of meanings in a larger range of daily life events, (ii) details a mechanism, the way to generate more signal possibilities within a constrained sound system. Whilst Nowak focused on (i), we focus on (ii). Nowak et al. expands on (i), that an additional selection pressure to increase combinations occurs when fitness benefits are gained from including more than one piece of information per utterance. This may be particularly so when combinations of meanings in an utterance carry more information than the production of each signal independently, such as detailing relational information between agents and events or agents and patients (e.g. Jackendoff & Wittenberg, 2017; Planer & Sterelny, 2021). Here, simply expanding a repertoire of single signals will not suffice. Franke (2016) demonstrates, by using game theory models, that compositional‐like structures can emerge generatively without requiring sophisticated cognition through spill‐over reinforcement learning, although this is disputed by Scott‐Philipps & Blythe (2013). Importantly, a simple combinatorial system, based on linear grammar that maps directly linear order onto semantic roles, as found in certain languages, does not necessitate complex cognitive abilities but can still denotate roles in an utterance such as agent, action and patient, allowing the communication a large array of messages (Jackendoff & Wittenberg, 2017; Truswell, 2017). Drawing on the social complexity hypothesis, we work with the assumption that as societies become more complex, species gain fitness benefits from communicating about more facets of daily life. Drawing from known or putative examples, these might include expressing concomitant information [e.g. the alarm + recruit vocal combination in Japanese tits (Suzuki et al., 2016), rest + feed vocal combinations in chimpanzees (Girard‐Buttoz et al., 2025)], or relations between individuals and events [e.g. signaller identity + travel/greet/feed vocal combinations in chimpanzees (Leroux et al., 2021; Girard‐Buttoz et al., 2022a, 2025)], which are not possible with single‐signal utterances (e.g. Bortolato et al., 2023a; Fig. 3). Accrued fitness benefits from utilizing signal combinations may result in further expansion of signal combination usage. Selection then acts to transform signal systems from a dependence on single signals to a dependence on signal combinations. Given the changing nature of how social or ecological events unfold in daily life, genetically fixed signalling systems involving minimal learning are likely to be too constrained to communicate liberally about daily life events. Hence, we expect some evidence that learning shapes the structure and usage of signal combinations. Empirical and theoretical animal examples illustrating four major transitions required to move from simple, non‐combinatorial to combinatorial communicators, across phylogeny or ontogeny. Given the same social scenario, whilst the left column illustrates signalling of a simple communicator, the right column illustrates animals who have reached each respective transition. Signals demonstrate increasing capacity to combine meanings about daily life events within utterances with each transition. We assume combinatorial communicators gain fitness benefits from using signals that encode meaning about numerous and concomitant daily events, leading to utterances which combine multiple meanings. The transitions are as follows: (i) signal combination diversity: moving from a reliance on single to combined meaning‐bearing signals; (ii) signal combination learning and conventionalization: social learning standardizes arbitrary vocal combination meanings within social groups or populations; (iii) diverse combinatorial mechanisms: usage of several combinatorial mechanisms expands expression of daily life events; and (iv) syntax: linear or hierarchical structuring of vocal combinations, e.g. referencing thematic roles: (i) first person, (ii) third person, and combining roles with daily events, e.g. ‘I greet you/him’. Changing the order or adding modifiers may change agent and patient roles. For each transition towards complexity, on the right, we represent empirical (pink speech bubbles) or theoretical (grey speech bubbles) examples of utterances individuals do/might make. Text boxes with yellow backgrounds indicate different combinatorial mechanisms. We use empirical examples for simple, non‐combinatorial communicators: baboons (Papio ursinus) (Transitions 1, 2 and 4) (Bergman et al., 2003; Maciej et al., 2013) and putty‐nosed monkeys (Cercopithecus nictitans) (Transition 3) (Arnold & Zuberbühler, 2008); and for complex communicators from Transitions 1–3: chimpanzees (Pan troglodytes) (Leroux et al., 2021; Girard‐Buttoz et al., 2025). For comparability, we draw on empirical examples from the primate literature but several transitions are also evident in other taxa (see Fig. 1 and main text for specific examples). Transition 4, right side, is theoretical, never having been described in a non‐human animal. Given a complex society, which includes most species represented in Fig. 2 (especially large multi‐male–multi‐female groups, cooperative breeding or fission–fusion societies), and drawing on signalling theories (see Section II), we assume (a) that there is selection pressure to generate more messages than there are single signal types. In line with Nowak et al.'s model, a generative combinatorial signalling system can only emerge if the same signal can be reused in numerous combinations. We expand on this idea by adding that four major mechanistic transitions are required to build such a signalling system (Fig. 3), in which all four transitions are combinatorial, and Transitions 3 and 4 may, but do not have to, include generativity. Whilst transitions 1 to 3 may overlap in their emergence, we expect that species which demonstrate transition 1 and 2 traits are the most likely to also demonstrate transition 3 and 4 traits. Throughout this review, where they exist, we provide empirical evidence of these transitions across the animal kingdom (summarized in Fig. 2). The first of these transitions is ‘signal combination diversity’ in which single meaning‐bearing signals are flexibly combined. The signal system shifts from a dependence on single signal utterances to reliance on utterances combining meaning‐bearing signals throughout daily life. In Transition 2, signal combination learning and conventionalization, usage learning of signal combinations and their meanings emerge. Learned arbitrary signal forms and functions must then adopt socially recognized conventions within communities for communication to be successful. In Transition 3, combinatorial mechanism diversity and versatility, signals that are emitted in short combinations (bigrams) using mechanisms that change or add signal meanings in different ways expand messaging, particularly by combining two meanings within utterances (Table 1; Fig. 1). Each mechanism can be used in several signal combinations using different rules or signals, possibly showing the emergence of a generative system. Versatility results in a system where messaging capacity shows expansion well beyond the available number of single signals such that everyday life events can be related and communicated. Combinatorial capacities are considered generative when they can encode information about numerous daily life events. The fourth and final transition is ‘syntax’. Combinatorial mechanisms become numerous and are exploited in longer signal combinations (trigrams or longer), containing at least three meanings within utterances. Hence, grammatical rules governing the organization of calls within utterances are required to convey relational information about daily events likely through a combination of linear or hierarchical structuring, which in turn facilitates message generativity. In sum, studies to date that have examined call combinations have mainly focused on alarm contexts (Fig. 1), such that the range of meanings at the scale of the full signal repertoire is not well documented. Likewise, the versatility with which such meaning‐bearing signals can be combined and recombined within utterances is largely unknown, with the exception of recent great ape and marmoset studies which demonstrate extensive versatility (Bosshard et al., 2024; Berthet et al., 2025; Girard‐Buttoz et al., 2025). Below, we detail the four major mechanistic transitions we propose, which we argue offer both evolutionary and ontogenetic routes from mainly rudimentary, single‐signal communicators to versatile, combinatorial, generative communicators. We detail how these transitions are measurable and predictable. Critically, we argue that simply having research on different species is not enough – research across species and ontogeny will be more powerful when directly comparable, using relatable, quantitative measures. SIGNAL COMBINATIONS The emphasis in this paper is on signal combination production as signal comprehension has been extensively covered elsewhere (e.g. Fitch & Friederici, 2012; Wilson et al., 2020). However, as signals produced can only be functionally relevant if they are comprehended, this literature is briefly summarized in below. Signal meaning For a communication system to evolve, signals must provide information that elicits receiver responses that benefit both signaller and receiver (Laidre & Johnstone, 2013; West et al., 2015). For instance, aggressive signals must generally work to repel, and copulation signals to attract, and coordination signals to coordinate. This is the case whether or not signals elicit mental representations or reflect the signallers' intentional goals, unintentional actions or simply the signaller's emotional state. In animal systems, determining the exact message encoded in an utterance is nonetheless challenging (Cheney & Seyfarth, 1990; Rendall, Owren & Ryan, 2009; Seyfarth & Cheney, 2017; Fischer & Price, 2017; Watson et al., 2022; Amphaeris et al., 2023) and has led to extensive interdisciplinary discussion about whether the information conveyed by animal signals to conspecifics constitutes meaning. Typically, repeated behavioural observations of natural signal production in specific contexts that lead to specific behavioural outcomes allow signal‐to‐event mapping needed to delineate species‐specific call types and their potential meaning or function (Seyfarth, Cheney & Marler 1980; Elie et al., 2024). The delineation of what constitutes a single call type and its relevance as an informative communicative signal in itself was then tested via carefully designed playback experiments that elicit relevant behavioural responses in the receiver in the absence of the associated context in which the call is usually elicited. Foundational studies ingeniously focused on alarm contexts to provide an external referent, the predator, facilitating discussions relating to mental representation and meaning (Seyfarth, Cheney & Marler, 1980; Fischer, 1998; Zuberbühler, Cheney & Seyfarth, 1999). Evidence of highly specific signal‐to‐context mapping across a range of species and contexts has led recent studies to conclude that meaning is an appropriate term when referring to the information contained in animal signals (Sievers & Gruber, 2016; Seyfarth & Cheney, 2017; Zuberbühler, 2019; Berthet et al., 2022; Watson et al., 2022; Amphaeris et al., 2023; Elie et al., 2024; Salis et al., 2025). We argue that examining shifts in contextual usage in call combinations compared to constituent calls offers an additional avenue to parse whether signallers are expressing contextual information or merely their emotional state. For example, chimpanzees emit screams when being attacked, a highly emotional context with negative valence. The other dominant context in which chimpanzees emit screams is in combination with a pant hoot vocalization during travel and food contexts (Girard‐Buttoz et al., 2025), resulting in contact or reunion with others, particularly allies (Mitani & Nishida, 1993; Crockford, 2019). Hence, the function of a scream may loosely relate to recruitment across contextual usage, but in this example, it cannot be underwritten by the same emotional or arousal state. Rather, contextual information appears dominant to emotional information. Here, the characterization of the call combination and its meaning is possible because of a fine‐grained knowledge of both signal form and function acquired over decades of research. Whether animal calls are analogous to human phonemes like [m] or [o], morphemes like ‘mon‐’ or ‘‐key’, words like ‘monkey’, and/or non‐verbal sounds like laughter (Anikin et al., 2023), remains an open question, largely because we do not know enough about how signal combinations operate across whole repertoires. However, the combinatorial scream example, suggests not all call elements are analogous to laughter, as laughter does not fundamentally change its contextual information when combined with other sounds or words. Signal intention Whether the signaller intends to change another's behaviour or mind through the use of a signal has also been a topic garnering intense research. Evidence, particularly in great apes, indicates that using signals with some voluntary control and intent to change behaviour is likely widespread in both vocal and gestural modalities of single signals (Great apes, Hobaiter & Byrne, 2011; Schel et al., 2013; Taylor et al., 2022; Bouchard & Zuberbühler, 2022; Monkeys, Silk, Seyfarth & Cheney, 2016; Cheney & Seyfarth, 2018). Hence, arguments that signals are only arousal‐driven likely miss crucial complexity. Signal use with apparent intent to change others' knowledge states has been argued for chimpanzees and bonobos (Pan paniscus) (Crockford et al., 2012; Crockford, Wittig & Zuberbühler, 2017; Schel et al., 2013; Genty & Zuberbühler, 2014; Moore, 2016; Girard‐Buttoz et al., 2020; Heintz & Scott‐Phillips, 2023; Townrow & Krupenye, 2025), giving insight into how animals think. Whether such signals are ostensive, however, remains untested. In contrast, human speech is considered to be ostensive, such that the signaller not only intends to change another's behaviour or mind but intends for the receiver to be explicitly aware of this intention (e.g. Moore, 2016; Graham et al., 2019; Heintz & Scott‐Phillips, 2023). It is assumed that non‐human animals do not engage in ostensive communication, partly because no test of ostension has ever been conducted, although test designs have been proposed (Moore, 2016). However, a new study testing and demonstrating metacognition in chimpanzees suggests meta‐awareness tests in the domain of communication may be feasible (Schleihauf et al., 2025). Comprehension of natural and artificial signal sequences Several species tested in captivity with artificial and meaningless sequences demonstrate understanding of both adjacent and non‐adjacent grammatical dependencies (see Petkov & Jarvis, 2012; Fitch & Friederici, 2012; Wilson et al., 2020, for reviews), suggesting some capacity to recognize simple non‐adjacent grammatical rules (see Table 1). Non‐adjacent grammatical rules that organize meaning in sentences are critical features of language Fitch & Friederici, 2012. However, this has been demonstrated neither in signal comprehension nor production in the natural communication system of any non‐human animal (Zuberbühler, 2020). However, relational meaning can be extracted from natural vocal sequences of two adjacent call types beyond the examples in Fig. 1, when the sequences are formed by more than one signaller. Playback experiments demonstrate that baboons keep track of the vocalizations of others (Crockford et al., 2007), interpreting them as third‐party interactions. For example, if baboons hear an aggressive threat call from one individual and a scream from another, baboons react more if the threatening individual is subordinate to the screamer (Bergman et al., 2003). This suggests that baboons extract a meaningful interaction from the two calls emitted by different signallers, here inferring a relationship between these two calls and callers that indicates that a rare rank change has occurred (Bergman et al., 2003; Cheney & Seyfarth, 2018). Given that baboons rarely emit combinatorial utterances (Boë et al., 2017; Hammerschmidt & Fischer, 2019) suggests that the proclivity to extract combined meaning from juxtaposed signals from different signallers is more common than the capacity to produce utterances that express combined meaning. In sum, such results have led to the conclusion that comprehension complexity outstrips production complexity, mirroring child language development. Critically, the assertion that hierarchical syntax is a human Rubicon remains (Seyfarth & Cheney, 2010). TRANSITION 1 – SIGNAL COMBINATION VERSATILITY: SEQUENCES SHIFT FROM CONTAINING SINGLE SIGNALS TO FLEXIBLY COMBINING MEANING‐BEARING SIGNALS If a species only uses one or two signal combinations when communicating, it suggests that any fitness benefits gained from combining signals do not generate sufficient selection pressure to promote the evolution of a generalized combinatorial communication system, which likely requires commensurate investment in relevant brain infrastructure. A particular challenge may be connecting motor and vocal brain regions with conceptual semantic regions, which, for the human language network, are in different lobes and connected by large white matter pathways (Gallardo et al., 2023; Fedorenko, Piantadosi & Gibson, 2024; Becker et al., 2025). In the absence of such selection pressures, a versatile production system that combines meanings and can detail an endless array of daily life events, including complex social scenarios, like speech does, seems unlikely to emerge. Studies initially assessed call combinations in alarm contexts. However, a range of species now also demonstrate meaning‐bearing calls are combined in other contexts (great apes, Leroux et al., 2021; Girard‐Buttoz et al., 2022a; Bortolato et al., 2023a; Berthet et al., 2025; monkeys, Bouchet et al., 2013; Coye et al., 2016; Bosshard et al., 2024; elephants, Pardo et al., 2019; meerkats, Rauber et al., 2020; and mongooses, Jansen et al., 2012). These examples across distantly related taxa demonstrate the capacity to convey up to two meanings per utterance. There is, however, as yet no conclusive demonstration that an animal species can convey three or more meanings per utterance, even if this possibility does not appear out of reach for species that recombine bigrams into trigrams and longer sequences like chimpanzees and marmosets (Girard‐Buttoz et al., 2022b; Bortolato et al., 2023a; Bosshard et al., 2024). The latter requires combinations of at least three meaning‐bearing calls or signals, for which there are also few documented examples across species (Bortolato et al., 2023a; Bosshard et al., 2024; reviewed for primate vocal sequences in Girard‐Buttoz et al., 2022b). This is clearly not due to articulatory constraints, given that song, which often requires rapid articulatory transitions (Elemans et al., 2008), is widespread across animal taxa, including in some primates. This leans to the possibility of cognitive constraints in semantic encoding, either at the level of comprehension of combined information units (see Section II) or at the production level of encoding meanings sequentially or hierarchically within a single utterance, points also raised by Planer & Sterelny (2021). A prerequisite of being able to flexibly combine meaning‐bearing calls is having the versatility to flexibly combine call types across the vocal repertoire. Studies that have examined whole repertoires at the species level typically find that 0–30% of utterances are emitted as sequences, and the diversity of sequences is limited given all possible sequences (see Girard‐Buttoz et al., 2022b for a review of primates; also see Fig. 5). To date, particularly great apes but also marmosets have been shown to flexibly combine calls into sequences and emit these sequences throughout a wide range of daily life events, not only in rare alarm contexts (Hedwig et al., 2014; Leroux et al., 2021; Girard‐Buttoz et al., 2022a, 2025; Bortolato et al., 2023a; Bosshard et al., 2024; Schamberg et al., 2024; Berthet et al., 2025). For example, each third utterance adult chimpanzees emit is a vocal sequence, and most call types combine with most other call types, rendering hundreds of versatile but structured sequences (Girard‐Buttoz et al., 2022b). Thus, vocal sequences are common, emitted throughout daily life during numerous events and social interactions. Of the 16 bigrams so far assessed to determine if sequences change the meaning of composite calls, 11 demonstrate clear meaning shifts. Assessing meaning content of repertoires with many sequences is a major task and is on‐going. However, repertoires with a diverse range of sequences across daily life are exactly the repertoires that are the most interesting for assessing the diversity of messaging. Theoretical sketch illustrating the accumulation of possible messages depending on the number of calls and combinations in a given vocal repertoire. We used empirical vocal data on sooty mangabeys (Cercocebus atys) (Sigmundson et al., 2025) and chimpanzees (Pan troglodytes) (Girard‐Buttoz et al., 2022b) to illustrate the number of different bigrams, trigrams and quadrigrams that have been reported so far in each species (with evidence of many more in chimpanzees than mangabeys). Each triangle (mangabey) and circle (chimpanzee) represent the number of combinations reported as a percentage of all possible theoretical bigrams (in blue), trigrams (in orange) and quadrigrams (in purple) that can be formed by recombining all the single calls in their repertoire in all possible orders. These percentages are projected on theoretical lines that represent the total number of messages that could be conveyed if each call combination were to convey a different meaning. Solid line: theoretical number of messages when ordering does not impact meaning, such that A_B and B_A convey the same message. Dashed line: theoretical number of messages if the order in which the calls are combined impacts meaning. The extent to which different vocal sequences convey different meanings remains to be shown for both species. Using not one but several of the combinatorial mechanisms illustrated in Fig. 1 increases the likelihood of reaching this maximum number of messages. In particular, ordering effects dramatically increase message generation potential, especially in longer sequences, although note that structured combinatorial systems such as human language are not expected to combine all sounds in all orders. We hypothesize that an early major transition towards a complex communication system is versatility in combining meaning‐bearing signals, enabling most signals to combine with most other signals throughout a variety of daily life events, and not only in rare events (e.g. alarm). Given that research determining signal meaning is time‐consuming, candidate species for combining meaning‐bearing signals can be detected by assessing species that commonly emit any signal sequences, whether meaning‐bearing or not, throughout numerous daily life events, and that sequence production is not limited to only one or two contexts. Such species would be prime targets to test for the subsequent major transitions required for complex communication detailed below. In contrast, animals that only emit sequences in few contexts can be excluded as candidates. Species can be quantitatively placed along a continuum, aiding both phylogenetic and ontogenetic comparison, by using simple metrics such as utterance length (number of different signals combined within one utterance) and diversity (number of different signal combinations) (see Fig. 4 for empirical examples). Comparative quantitative measures of combinatorial capacities highlighting variation across species and ontogeny. The networks on the left represent all vocal bigrams (two‐call combinations) documented between single call types at the scale of the full repertoire in adult sooty mangabeys (Cercocebus atys; above) and chimpanzees (Pan troglodytes; below), data published in (Girard‐Buttoz et al., 2022b; Sigmundson et al., 2025; Le Floch et al., 2026a). These networks illustrate higher call combination diversity in chimpanzees compared to sooty mangabeys. Orange circles depict single call types, and blue arrows show how calls are combined in bigrams. The width of the arrow is proportional to the number of occurrences of each bigram. The right panels, ontogenetic combinational change, illustrate schematically the changes in vocal sequence length (orange) and diversity (green) across ontogeny in those two species [schematic curves redrawn from empirical data published in Bortolato et al. (2023b) and Sigmundson et al. (2025)]. The individual developmental stage is illustrated with a black silhouette and its mother with a grey silhouette. Whereas both combinatorial metrics remain relatively stable across the lifespan in mangabeys, chimpanzees experience a dramatic increase in both vocal sequence length and diversity across the first decade of life. This comparative example illustrates the value of quantitative whole repertoire approaches for phylogenetic and ontogenetic comparisons. TRANSITION 2 – SIGNAL COMBINATION LEARNING AND CONVENTIONALIZATION: SOCIAL LEARNING STANDARDIZES NON‐INNATE COMBINATIONS AND THEIR MEANINGS WITHIN COMMUNITIES Most animal signals are largely genetically fixed (reviewed in Fedurek & Slocombe, 2011; Fischer & Hammerschmidt, 2020) in both their form (the shape of the signal) and their function (the usage of the signal and the outcome it produces) (Table 1). As such, individuals of a given species will emit the same or very similar signal forms to similar stimuli, even in geographically or genetically distant populations, excluding the possibility of social learning (Mitani, Hunley & Murdoch, 1999; Catchpole & Slater, 2003; Crockford et al., 2004, p. 2018; Fischer & Price, 2017; Hammerschmidt & Fischer, 2019). Much rarer are signals or their combinations that demonstrate population‐level meaning shifts, showing an apparent ‘decoupling of the direct, one‐to‐one linkage between meaning and form’ (Dingemanse et al., 2015, p. 610; Watson et al., 2022). Such a decoupling suggests an element of learning, where learning can then lead to an arbitrary connection between a signal and its meaning. To become communicatively useful, signal‐to‐meaning mappings need to be shared across the community of communicators (e.g. family, region, country) through social learning, a process referred to as conventionalization (Lewis, 1969b; Moore, 2013). In an arbitrary system, such as human language, the signal form is a symbolic representation of the semantic content it relates to. Learning occurs for both the signal form (‘ape’ takes the form ‘Menschenaffe’ shifting from English to German), and the signal function, contextual usage or semantic content (‘ape’ may change from referring to a particular animal to the action of imitating an ape: ‘to ape’). Given that only humans are thought to have an arbitrary signalling system, the challenge of how species may shift from innate to arbitrary signalling remains empirically unresolved. In linguistics, signal conventionalization is defined as requiring signaller insight into any community‐wide agreement on a particular signal meaning (e.g. Lewis, 1969b). Given early child language‐learning, Moore (2013) argues more simply that signallers are only required to learn that a signal can achieve a particular function, and to imitate the form and usage of the signal. Here, neither insight nor meta‐representation are required to achieve conventionalized signalling through conformity. Conventionalization does not require vocal production learning; it can emerge through social usage learning. Below, we explain why – unlike vocal usage learning – vocal production learning offers limited potential for acquiring new meanings. We then explore how vocal usage learning, especially the learning of signal combinations, can facilitate meaning change and conventionalization. Finally, we emphasize the importance of tracking signal acquisition and usage across development to assess the potential for vocal usage learning. Vocal production learning rarely leads to changes in meaning In the vocal domain, generating entirely or relatively new signals requires vocal production learning (Janik & Slater, 2000; Vernes et al., 2021). Vocal learners, like parrots, songbirds, whales and dolphins, are adept at this. In their natural communication system, their strong vocal production learning capacities seem to function to generate group or individual identifiers that mostly serve three functions (within group identification, territorial defense or mate attraction (Janik & Slater, 2000; Catchpole & Slater, 2003; Bradbury & Balsby, 2016); but see Loning, Griffith & Naguib, 2024a). As such, vocal production learning capacities rarely lead to the acquisition of new meanings (Petkov & Jarvis, 2012; Loning, Griffith & Naguib, 2024b). In some animals (elephants, Pardo et al., 2024; marmosets, Oren et al., 2024; dolphins, King & Janik, 2013) signallers adjust the acoustic properties of their signals to the recipient they address, but the function of the call type itself (an identifier) does not change. In sum, whilst vocal production learning has undoubtedly evolved many times, cases are rare where such learning leads to an expanded capacity to communicate meaning (Berwick et al., 2011; Petkov & Jarvis, 2012; Jarvis, 2019). Future research on vocal sequences in these species may yet demonstrate otherwise. Usage learning may more readily lead to capacity for meaning change and conventionalization Changing the signal function or meaning requires vocal usage learning, the capacity to emit an existing sound in a different or new context or to learn to combine signals in sequence (Janik & Slater, 2000; Vernes et al., 2021). Identifying the use of arbitrary signalling would require identifying a vocalization that is given a new meaning, which bears little or no intrinsic resemblance to the original meaning (Lewis, 1969b; Moore, 2013; Planer & Kalkman, 2021; Gasparri et al., 2023), for example, by observing population or ontogenetic differences in usage. In humans, whilst both vocal production and usage learning are required to become proficient speakers (Vernes et al., 2021), given that all human languages construct words from a limited pool of 150 possible phonemes (Moran et al., 2012), the question of the extent to which speech actually relies on vocal production learning rather than vocal usage learning is raised. Whilst production learning requires novel sound generation, usage learning requires novel ways to use existing sounds, such as through combinatorial processes. Usage learning examples to date observed in natural animal systems usually involve a limited contextual shift of single signal use, such as monkey predator alarm calls being elicited by a narrower range of predator types through ontogeny (Seyfarth & Cheney, 1980; León et al., 2023) or by novel object presentations with overlapping properties to those of the predator, such as drones eliciting aerial predator alarm calls (Wegdell, Hammerschmidt & Fischer, 2019). Exceptions are the ‘language‐trained’ apes and birds, which produce sequences using human‐made arbitrary signalling systems (bonobos, Savage‐Rumbaugh et al., 1993; gorillas, Tanner & Perlman, 2017; African grey parrots (Psittacus erithacus), Pepperberg, 2021). In general, once individuals can flexibly combine their calls to create novel sequences, they can either emit novel idiosyncratic sequences, produced only by themselves or copy the sequences of others. Meaning then needs to be mapped onto novel sequences through vocal usage learning processes (Vernes et al., 2021). Examples evoking usage learning involving signal combinations in natural animal repertoires are evident in great apes. A bonobo call combination indicates a shift in function, where the bigram whistle‐hoot shows a contextual shift across two wild populations (Schamberg et al., 2024). Also, a wild chimpanzee call combination shows population shifts in call order. Here, the bigram pant hoot + pant grunt emitted in fusion + greeting contexts is produced as pant hoot first in a population of western chimpanzees and as pant‐grunt first in an eastern chimpanzee community (Girard‐Buttoz et al., 2022a). The contextual usage is consistent across populations but the respective usage of call order has become standard within populations. Finally, in zoo chimpanzees, an oral‐articulated ‘raspberry’ sound can be added de novo into a vocal sequence and then adopted by community members (Marshall, Wrangham & Arcadi, 1999), also suggesting propagation through social learning. Of note, in the gestural modality, great apes also show shifts in both form and function, albeit only documented in single signals (Badihi et al., 2023a; Malherbe et al., 2025; reviewed in Kalan, Nakano & Warshawski, 2025). A number of studies also point to a degree of voluntary control, especially in great ape gestural and vocal signalling (see Section II). Whether this variety of examples indicating usage learning suggests some shift in communicative capacities in hominids or simply under‐reporting in other species requires further research. Ontogenetic assessment of signal learning Across ontogeny, signal combination learning capacity can be assessed by comparing the length and diversity of signal combinations with age (Fig. 4), as well as signal ordering and usage. Learning capacity can be ruled out if infants emit the full repertoire of adults. For single signals, primate cross‐fostering studies between closely related macaque species have demonstrated strong call innateness, as infants only emitted calls of their genetic rather than foster species (Owren et al., 1993). To assess whether vocal sequence production is similarly innate, studies have assessed vocal sequence change across ontogeny. Sooty mangabeys (Cercocebus atys), a catarrhine monkey, emit over 40 different sequences, but only three are commonly used across individuals (Le Floch et al., 2026, 2026). Critically, adults do not have longer sequences, nor a greater diversity of vocal sequences than their infants (Sigmundson et al., 2025; Fig. 4), suggesting limited scope for vocal sequence learning. The common marmoset (Callithrix jacchus), a platyrrhine monkey that utters a large diversity of vocal sequences across contexts which shift with age, in contrast, shows potential for learning the usage of certain vocal sequences. However, whilst sequence diversity increases in subadults, it shows similarities in infant and adult phases (Gultekin et al., 2021). Potential for usage learning of vocal sequences seems higher in chimpanzees, where vocal repertoires expand three‐fold across the first decade of life, shifting from rudimentary utterances of mostly single calls to widespread use of a diverse range of vocal sequences as adults (Bortolato et al., 2023b; Fig. 4). Whilst these patterns may be explained by maturational processes, when viewed together with a chimpanzee study showing population variation in vocal sequence structure (Girard‐Buttoz et al., 2022a), they suggest social learning may be explanatory for at least some chimpanzee vocal sequences. In sum, examples in the natural communication systems of great apes suggest that there is sufficient flexibility in signal form and function to warrant testing for the existence of conventionalization processes. This raises two possibilities. First, conventionalization processes arise even with extremely limited usage flexibility such that most of the repertoire remains fixed, or second, there is more usage flexibility in these systems than has been detected to date. Nonetheless, no species other than humans is known to use arbitrary signals with conventionalized signal meanings (Skyrms, 2010; Planer & Sterelny, 2021). It may well be that switching to a conventional system creates demands on pragmatic interpretation, as conventions are likely harder to interpret, generating barriers to their evolution (Aussems & Moore, 2025). We hypothesize that species that demonstrate both the use of a large diversity of signal combinations (Transition 1) and the decoupling of single signal form and function (Transition 2) are the most likely to demonstrate conventionalization of signal combinations at the level of the social group. We predict that species which use call combinations that particularly demonstrate shifts in function across populations are prime candidates. We do not expect vocal production learning capacities to be influential with respect to a species capacity to be a vocal usage learner. This is because adept compared to more limited vocal production learners do not show greater capacity to embed meaning into utterances, nor in vocal usage learning. TRANSITION 3 – COMBINATORIAL MECHANISM DIVERSITY AND VERSATILITY: USAGE OF SEVERAL COMBINATORIAL MECHANISMS EXPANDS THE EXPRESSION AND COMBINATION OF DAILY EVENTS Once a system is in place that can flexibly combine sounds, strong selection pressure to expand the number of messages that can be conveyed is likely required to shift signalling systems to liberally encode combined meanings. It seems unlikely that such a system would emerge to expand messages only about rare events, such as alarm events. More likely, such an emergence would occur when conveying messages that combine information about numerous daily events becomes advantageous (Nowak & Krakauer, 1999). We argue that a feature particularly likely to promote investment in a combinatorial system is the need to convey concomitant information about everyday events, such that more than one meaning is needed per utterance, such as to encode relational information between actors and/or events. This requires a signalling system (i) that liberally combines at least two meaning‐bearing signals within an utterance, (ii) where the signal form or function are not necessarily innate but may be learned, (iii) leading to short utterances which routinely convey more than one piece of information. Several bird species [pied babblers (Turcoides bicolor) and at least one species of tit (Parus minor), Suzuki et al., 2016; Engesser et al., 2016; Dutour et al., 2019] each produce a structured call combination that signals about two synchronous contexts/events – alarm and recruitment (Fig. 1). Whilst impressive, there is only one example per animal repertoire related to a single external stimulus (the presence of a predator). However, there are many occasions throughout an animal's daily life when synchronous events occur, such as greeting an individual in a food patch (Fig. 3). Language can effortlessly incorporate such juxtapositions of events into an utterance e.g. “Hi, let's eat”. But can other animals? We argue that when faced with two synchronous events, a combinatorial communicator will communicate about both in one utterance, whereas a simple, non‐combinatorial communicator will only communicate one event, such as greeting rather than food. However, for most species, at most only a few vocal sequences have been reported so far (reviewed in Engesser & Townsend, 2019; Girard‐Buttoz et al., 2022b). This precludes even the possibility that utterances can be used liberally to inform others about a range of synchronous everyday events. However, chimpanzees may liberally use vocal sequences to convey information about concomitant events (Bortolato et al., 2023a). Chimpanzees were more likely to emit utterances containing single calls when the signaller encountered only one event at a time (e.g. only feed, greet or aggress). The likelihood of uttering vocal sequences doubled when chimpanzees were engaged in two or more events that occurred concomitantly, such as greeting while feeding (Fig. 3). This indicates not only the structural possibility to liberally combine information within sequences but also a potential to liberally combine meanings detailing synchronous everyday events (Bortolato et al., 2023a). Indeed, chimpanzees are the first non‐human species to demonstrate the use of not one but all combinatorial mechanisms currently identified in animals (Fig. 1), such that call combinations can either modify, add and generate meanings (Girard‐Buttoz et al., 2025) and that meaning is altered by the order of calls in the combination. Furthermore, each mechanism is used for more than one call combination and contextual usage, demonstrating versatility in the usage of combinatorial mechanisms across call combinations. Furthermore, the same call type is used for more than one mechanism. The combined capacity to generate many call combinations across daily life and utilize several combinatorial mechanisms likely exponentially increases the capabilities to generate messages, in particular when the order in which the signals are being combined alters the meaning of the combination (Fig. 5). Similar versatile use of call combinations across contexts has also been reported in wild bonobos (Berthet et al., 2025) but the combinatorial mechanisms involved remain debated (see the accompanying eletter by Chemla et al. (2025) “The plus and minus of bonobo call combinations”). The observed versatility in signal combination usage in chimpanzees is unexpected. It demonstrates that a signalling system with mainly inherited single signals, since chimpanzees across populations produce very similar calls (Crockford, 2019), may engage versatility when combining signals. Indeed, signal combinations may show a range of changes in signal usage and function that are not evident for single calls. This demonstrates that the capacity for usage learning, whereby existing signals can be lifted out of one context and used in another, may be critical for generating complex combinatorial communication systems. The extent to which combinatorial flexibility is possible with limited vocal production learning remains to be assessed but may be much more extensive than previously considered possible. Likewise, the extent to which vocal production learning in vocal learning species leads to contextual change also needs to be assessed (see Transition 2 above). To date, combinatorial mechanisms identified in animal communication mainly refine meanings rather than add meanings together. Indeed, examples of the latter remain rare (Fig. 1). Nonetheless, utterances need to contain more than one meaning in order to express relations between objects and events, a basic function of language. As more combinatorial mechanisms are utilized by a species, the more they achieve versatility in messaging. Thus, we hypothesize that species that use, not one, but various combinatorial mechanisms to expand meanings relating to daily life will be more likely to emit short utterances containing at least two meanings, preluding the emergence of syntax. Whilst Transitions 1–3 may emerge in any and overlapping order, we expect species that have achieved Transition 1 and 2 will be more likely to demonstrate capacities required for Transition 3. Specifically, we expect a certain versatility such that meaning‐bearing calls can be liberally combined with each other to communicate about changing juxtapositions of daily events. We expect versatility in signal usage will require learning processes, possibly including conventionalization. Once capacities needed for Transitions 1–3 are in place (i.e. once there is flexibility in signal sequence usage, structure, structural rules, and conventionalization of meaning) what constraints might be in place to stop a signalling system that incorporates meaning from rapidly and exponentially expanding? We list three additional considerations that might prevent further expansion. First, if few signals in the repertoire are under voluntary control, they cannot be liberally combined. Second, there may be limited numbers of messages to convey, possibly limited by conceptual thought. For example, signals that convey information pertaining to displacement in time and place, have only been reported in one non‐human species, the orangutan (Pongo abelli) (Lameira & Call, 2018). In particular, whilst some mammals and birds are capable of planning future actions and have long‐term memories (Baddeley et al., 2001; Boeckle & Bugnyar, 2012; Bruck, 2013; Janmaat et al., 2014; Keenan et al., 2016), they may rarely communicate about past or future events. Third, there may be limited capacity to encode signals into hierarchical structures using rules that could facilitate meaning extraction from many sequences. The capacity for hierarchically structured thought is considered to constitute a major evolutionary milestone in cognition, impacting both abilities for technological and communicative complexity (Friederici, 2011; Planer, 2023), which we return to in Transition (4). TRANSITION 4 – SYNTAX: INTRODUCING LINEAR OR HIERARCHICAL STRUCTURING INTO SIGNAL COMBINATIONS Human utterances are typically relational, relating different information within an utterance to offer disambiguation, nuance and interpretation about an event or object of interest (Skyrms, 2010; Planer & Sterelny, 2021; Zuberbühler & Bickel, 2022). Whilst relational information minimally requires linear processing (Jackendoff & Wittenberg, 2017), syntactical structures like embedded clauses require hierarchical organization (e.g. Skyrms, 2010; Planer & Sterelny, 2021; Zuberbühler & Bickel, 2022). Language‐trained apes have demonstrated the use and comprehension of linear symbol constructions in utterances such as ‘Kanzi ball’, referring to self and an item self was interested in having (Truswell, 2017; Berio & Moore, 2023). For decades, such capacities have been considered to be due to enculturation processes derived from human exposure involving exaptation (e.g. Lloyd, 2004). However, recent analyses suggest natural animal communication systems may have similar constructs by combining calls that strongly encode signaller identity together with calls that indicate an event, such as signaller identify + food or greet in chimpanzees (Leroux et al., 2021; Girard‐Buttoz et al., 2022a). Further, several other compositional‐like vocal utterances have recently been identified in natural Pan communication (Berthet et al., 2025; Girard‐Buttoz et al., 2025), raising the question of whether language‐trained ape skills rather reflect natural communication capacities. Hierarchical structuring of meaning‐bearing units within utterances, on the other hand, is deemed the exclusive realm of human language (Chomsky, 1965; Hauser et al., 2002; Fitch & Friederici, 2012; Jackendoff & Wittenberg, 2017; but see Frank, Bod & Christiansen, 2012). Whilst natural animal communication examples of ordering effects on the meaning of call combinations have been interpreted as encoding syntactical relations, to date, these examples all relate to one or few combinations involving a minority of signals in the repertoire, combining a maximum of two meaning‐bearing units. They detail either short morpho‐syntactical structures, where one unit modifies the meaning of another (Ouattara, Lemasson & Zuberbühler, 2009; Girard‐Buttoz et al., 2025) or where the meanings of each constituent call are conjoined (Suzuki et al., 2016; Engesser et al., 2016; Girard‐Buttoz et al., 2025). Such examples are limited to bigrams, offering possible cases of compositionality or proto‐syntax (Suzuki & Zuberbühler, 2019; Planer & Sterelny, 2021; Berthet et al., 2022; Schlenker et al., 2023). In contrast, hierarchical structuring, which requires recombining and embedding of shorter structures (like bigrams) in longer structures can only appear in combinations containing at least 3 units. As modelled by Nowak & Krakauer (1999), for a combinatorial system to emerge, first, the total number of messages must likely exceed a critical value well beyond the number of single calls, and second, individual calls should occur in many different messages. For syntax to emerge in communication, we suggest that a critical mass of messages needs to be conveyed with a critical mass of utterances containing at least 3 units. Once selection pressures are sufficient, a species may invest in the cognitive hardware to generate a MERGE‐like hierarchical system (Bolhuis et al., 2014; Rizzi, 2016), enabling recombination or recursion of short chunks (of two meaning‐bearing items) within a phrase, for example embedding the short chunk [the ape] into a longer utterance [I like the ape]. The idea of animals using signal chunks which can be embedded into longer signal sequences evokes such MERGE‐like structures (Zaccarella, Schell & Friederici, 2017; Suzuki & Zuberbühler, 2019; Suzuki & Matsumoto, 2022). Whilst no non‐human species has unequivocally demonstrated the production of relevant meaning‐bearing structures, interestingly, two primate species, chimpanzees and marmosets, demonstrate the capacity to embed independently emitted bigrams into longer vocal sequences (Girard‐Buttoz et al., 2022b; Bosshard et al., 2024). In chimpanzees, for example, 84% of bigrams (49 out of 58) could have a third call type added to them to produce trigrams. As well as phrase embedding, another key function of syntax is the denoting of thematic roles of actors and actions, a near universal combinatorial utterance structure in human language (Townsend et al., 2018; Zuberbühler & Bickel, 2022). Simple hierarchical structured sentences most commonly contain thematic relationships of agent, action, and patient in various language‐dependent orders. Such sentences typically denote an actor or agent, who does something (action, verb) to someone or something (the patient or object). The syntactical demarcation of the thematic roles means that in each case, agents, actions and patients can be switched out and effortlessly replaced by different agents, actions and patients. This, combined with recursion, facilitates the open‐ended capacity of language to infinitely generate new utterances. It has been argued that the lack of open‐endedness in animal communication in itself indicates a lack of syntax (Bolhuis et al., 2018). Palazzolo (2025) alternatively argues for a more continuous evolutionary scenario, that the first hierarchical structures were likely limited in the number of units they could embed. Many meanings identified as being conveyed by animal signals seem to relate to observable actions (e.g. groom, rest, play, travel, copulate). Some relate to objects (e.g. predator, food), some encode the signaller's identity (e.g. contact calls in dolphins, marmosets, meerkats, elephants, and chimpanzees), the identity of the receiver targeted by the vocalization (in dolphins, King & Janik, 2013; elephants, Pardo et al., 2024; and marmosets, Oren et al., 2024), or combinations of these (see Fig. 1). Some signal functions also relate to non‐observable actions, where the apparently desired action is not yet visible as the signal precipitates the action (e.g. recruitment signals may precipitate the recruitment of others). However, roles can only be demarcated in relation to each other, and hence can only begin to become relevant in utterances which contain at least two meanings. Indeed, animal signal combinations which have demonstrated at least two meanings within an utterance are few. Whilst it remains to be determined whether such combinations depict thematic roles in any animal other than humans (see Truswell, 2017; Berio & Moore, 2023), there are several interesting candidates to test. We suggest how these examples may relate to thematic roles. Some may combine action and patient (e.g. Japanese tits: recruit + predator, Suzuki et al., 2016; bonobos: recruit + travel, Schamberg et al., 2016), or agent and action (e.g. meerkat and banded mongoose: signaller Identity (agent) + action, Jansen et al., 2012; Rauber et al., 2020; chimpanzee: signaller identity (agent) + greet, Girard‐Buttoz et al., 2022a), or signaller identity (agent) + patient (chimpanzee: signaller identity + food, Leroux et al., 2021; bonobo: action + signaller identity, Schamberg et al., 2016). Particularly the meerkat, mongoose, chimpanzee and bonobo combinations may even point towards animal signal productivity, where a signal retains its meaning or role across combinations (Berthet et al., 2022). Importantly, the semantic category of each signal (agent, action, patient) can emerge without hierarchical syntax simply using a linear grammar. Here, semantic roles can be identified by combining linear ordering of signals and associated pragmatic information (Jackendoff & Wittenberg, 2017; Truswell, 2017; Aussems & Moore, 2025). Such a communication system is not beyond the reach of animals and can be sufficiently generative even for human communication, since some human languages may rely largely on linear grammar (Jackendoff & Wittenberg, 2017). We hypothesise, however, that hierarchical grammar can be seen as a complexification of linear grammar and likely emerged later in evolutionary time, such that hierarchical structures may build on linear structures. Recent studies demonstrate that, conceptually, primates when watching video sequences of animate or inanimate agents and patient interactions effortlessly and appropriately encode agents (actors) and patients (receivers) (Zuberbühler & Bickel, 2022). Taken together, this raises the possibility that vocal sequence structures that combine signals which primarily encode signaller identity with signals that encode actions, like greet, may encode thematic roles like agent + action, such as ‘I greet’, which are also suggested as possible examples of predication (Collier et al., 2014; Leroux et al., 2021). However, these ideas currently remain untested. Nonetheless, the sense of potential continuity between animal and human communication about events in the real world has given rise to the event perception hypothesis of the evolution of syntax, that the origins of syntax lie in how signallers perceive and mentally represent events (Zuberbühler, 2019, 2020), however tests of this in animal systems are required. Given this apparent palette of animal signal meaning combinations within utterances, in the search for hierarchical structure in animal utterances, consideration of roles of agent, patient, action, and their productivity, might not be out of reach. Whilst potentially difficult to test using conventional playback experiments, the apparent ease with which chimpanzees assign agency to avatars in virtual worlds (e.g. Allritz et al., 2022; Rapport Munro et al., 2025b) opens the possibility to test comprehension of conspecific vocal sequences in virtual world contexts. We hypothesize that animals that routinely combine more than two meanings within utterances, recombine signal chunks into longer sequences, and use signals with potential to convey different thematic roles, are prime candidates for utilizing syntax and having generative capacities in utterance construction. Species with Transitions 1–3 would be candidates for examining Transition 4. FUTURE DIRECTIONS AND CHALLENGES Accurate mapping of signal complexity in natural animal systems is relevant across a rich cross‐disciplinary field, from evolutionary models of language evolution (Chomsky, 1965; Nowak & Krakauer, 1999; Nowak & Komarova, 2001; Collier et al., 2014; Bolhuis et al., 2014; Zuberbühler, 2020; Planer & Sterelny, 2021) to comparative models of gene and brain function (Enard et al., 2002; Gallardo et al., 2023; Matentzoglu et al., 2025; Becker et al., 2025), all of which will remain limited until we crack animal signal code to discover its complexity. Natural observations of animals communicating in their natural socio‐ecological environment are an imperative starting point. This determines how and when animals use each signal in their repertoires. More specifically, only natural observation can identify contexts of signal usage and targeted audiences, whether close or distant. Many contexts that trigger signals under natural socio‐ecological conditions are not present in captive settings and cannot be guaranteed in wild settings with human disturbance. For example, chimpanzees emit context‐specific signals in contexts that do not necessarily occur in disturbed habitats or captivity, such as when hunting monkeys or soliciting copulations. In environments with no hunting possibilities or no sexually receptive females, the context‐specific ‘hunt bark’ signal (Mitani & Watts, 1999; Crockford & Boesch, 2003; Mine et al., 2022) or palette of copulation solicitation signals (Badihi et al., 2023b; Malherbe et al., 2025) will be absent in the chimpanzee repertoire. It is unclear how many vocal sequences might only be emitted in wild settings. Captive settings, however, offer the possibility to assess the impact of exposure to novel contexts on signalling. For example, in captive settings, interactions with humans can elicit novel vocal (African grey parrots, Pepperberg, 2006; orangutan, Lameira et al., 2013) and gestural (chimpanzees, Leavens, Hopkins & Bard, 1996; gorilla, Tanner & Perlman, 2017) signals and signal sequences (bonobo, Savage‐Rumbaugh et al., 1993; parrot, Pepperberg, 2021), creating new signal forms and functions. The extent to which signals demonstrate new form and function may give insights into the extent to which animal signal repertoires are fixed or flexible. Captive settings also offer a new avenue in deciphering animal sequences with touch‐screen or avatar experiments (Rapport Munro et al., 2025a), allowing individuals to actively choose meanings related to particular signals and signal combinations. Hence, a strong approach to decoding animal signal sequences is to compare signal sequence usage across wild and captive populations. Animal communication research has been effusive, but the capacity to make rapid progress in deciphering animal signal sequences has been limited by highly time‐consuming tools such as field playback experiments or conducting acoustic analyses, which for noisy signals in noisy environments can still require coding elements within sequences by hand. The potential of machine learning to code and classify vocal repertoires into call types assigned to particular contexts will likely have a major impact on progress (Rutz et al., 2023). Whilst automated coding and classifying has vastly increased speed of processing and analyses for species with mainly tonal calls in discrete call systems, such as bird song (Xie et al., 2023), it remains a significant challenge for graded call systems containing a mix of both tonal and non‐tonal calls, such as is the case for great ape vocal repertoires. At least some progress has been made (Fedurek, Zuberbühler & Dahl, 2016; Arnaud et al., 2023; Valet et al., 2025) and sizeable vocal recording libraries now exist for training machine learning software (e.g. Girard‐Buttoz et al., 2022b; Bortolato et al., 2023b). Surprising, given previous conceptions of fixed, innate vocal repertoires in non‐human primates, is that our closest living relatives are emerging as a front‐runner for vocal communicative complexity. There is versatility in both vocal and gestural sequence construction in the natural communication systems of gorillas, bonobos, and chimpanzees, with at least bonobo and chimpanzee sequences including meaning‐bearing signals (Transition 1). Chimpanzees like humans also demonstrate a dramatic ontogenetic shift across the first decade of life, with infants being rudimentary single signals users and adults relying on combinatorial communication. There is flexibility in signal combination form and function in their natural communication systems, indicating capacity for some social learning of vocal and gestural signals (Transition 2). Bonobos and chimpanzees demonstrate a number of meaning‐bearing call combinations, in which meaning is altered or added compared to the composing units, with chimpanzees and bonobos doing so using several combinatorial mechanisms which flexibility engages different call types (Transition 3). Taken together, great apes demonstrate the most generalized combinatorial capacities, including most of those required to build a complex, combinatorial communication system. Whether these systems indicate a particular shift in the hominid lineage towards becoming versatile, generative communicators requires further research. To date, other species which are good candidates for testing for combinatorial communication systems, because they at least show several signal combinations in different daily life contexts, are also often those that live in more complex societies particularly cooperative or fission–fusion social systems (Figs 1 and 2), evoking the social complexity hypothesis. Nonetheless, to date, identified great ape capacities place them as highly likely candidates to be combinatorial, generative communicators. Given progress in understanding animal signalling relies heavily on observing animals in their natural habitats, and because native great ape habitats are under huge threats, if we are to ever crack the code, both research and conservation efforts are urgent. CONCLUSIONS (1)Accurate quantitative assessment of the combinatorial capacity of natural animal communication systems is required for comparative and evolutionary models, such as for brain function, genes or language. Accurate quantitative assessment of the combinatorial capacity of natural animal communication systems is required for comparative and evolutionary models, such as for brain function, genes or language. (2)Recent animal studies point to a clear empirical direction to decode the combinatorial strategies that animals use in their natural communication. To determine which species are predominantly combinatorial signal users rather than non‐combinatorial signal users, we offer an utterance evolution framework which details four major mechanistic transitions required to build a generative combinatorial system, which may show similarities and differences to language, and which offers quantitative measures for phylogenetic and ontogenetic comparison. Recent animal studies point to a clear empirical direction to decode the combinatorial strategies that animals use in their natural communication. To determine which species are predominantly combinatorial signal users rather than non‐combinatorial signal users, we offer an utterance evolution framework which details four major mechanistic transitions required to build a generative combinatorial system, which may show similarities and differences to language, and which offers quantitative measures for phylogenetic and ontogenetic comparison. (3)Whilst a wide range of species use signal combinations to modify, add or generate meanings, signal combinations and combinatorial mechanisms are limited per species. Whilst a wide range of species use signal combinations to modify, add or generate meanings, signal combinations and combinatorial mechanisms are limited per species. (4)To date, the extant hominids demonstrate the most generalized and versatile combinatorial capacities, including many required to build a combinatorial communication system. To date, the extant hominids demonstrate the most generalized and versatile combinatorial capacities, including many required to build a combinatorial communication system. (5)Further research is needed to detail whether the signal combinatorial skills in great ape communication demonstrate that a shift towards language is already evident in our extant hominid relatives, or rather reflects under‐reporting in non‐ape species. Further research is needed to detail whether the signal combinatorial skills in great ape communication demonstrate that a shift towards language is already evident in our extant hominid relatives, or rather reflects under‐reporting in non‐ape species. (6)In sum, we argue that recent research showing diverse combinatorial capacities across a wide range of taxa and versatile combinatorial capacities in great apes, means that it is no longer parsimonious to assume that humans are the only species with a generative communication system. Rather, this requires empirical testing. In sum, we argue that recent research showing diverse combinatorial capacities across a wide range of taxa and versatile combinatorial capacities in great apes, means that it is no longer parsimonious to assume that humans are the only species with a generative communication system. Rather, this requires empirical testing. ACKNOWLEDGEMENTS We are forever indebted to the late Christophe Boesch, whose passion for chimpanzees and their societies has and will always, richly inspire and influence our thinking. We thank Derry Taylor, Robert Seyfarth, Nicolas Mathevon, Richard Moore and two anonymous reviewers for highly instructive comments on earlier versions of this work. This work was supported by the Max Planck Society (M.IF.EVAN8103 – to C.C. and R.M.W. through the Evolution of Brain Connectivity Project). The funders had no role in this review, decision to publish, or preparation of the manuscript. Open access publication funding provided by COUPERIN CY26. DATA AVAILABILITY STATEMENT Data sharing not applicable to this article as no datasets were generated or analysed during the current study. REFERENCES Associated Data Data Availability Statement Data sharing not applicable to this article as no datasets were generated or analysed during the current study.