The “canonical” list of vocal learners

There is a general trend in the literature (e.g., [8]) that limits vocal learning only to species that can produce novel calls through imitation, subserved by a direct connection between the forebrain and phonatory muscles (e.g., the larynx in mammals or syrinx in birds). This has led to a canonical list of vocal learning species. It comprises three bird orders and some mammals. The birds—by far the longest list when counting individual species (in the thousands [9])—comprise songbirds (Passeriformes), parrots (Psittaciformes), and hummingbirds (Trochiliformes) [10, 11]. The mammals include humans, some cetaceans [12, 13], pinnipeds [14, 15], elephants [16], and bats [17, 18]. Birds are considered closest to humans in vocal learning abilities, even though they are phylogenetically the most remote. Humans are the only primate uncontroversially considered to be vocal learners, whereas nonhuman primates are considered of little relevance in this regard. Refinements to this list usually consist of looking inward for finer distinctions in the families already established (e.g., bats [18] or parrots [19]), and rarely outward.

It is possible that this focus on imitation and novel vocalizations is due to it being the clearest case of something being “transmitted” and then “learned.” That is, it could be that for those who put a premium on imitation (e.g., [8]), learned (as opposed to “innate”) entails that there has to be imitation of something that wasn't there before in any form (in the repertoire), as opposed to improvisation or other ways in which sounds in a repertoire can change (for example, through social feedback or modifications of aspects of calls that do not entail an entirely novel output).

It is also possible that the appeal of “neuro-reductionism” (to virtually equate a behavioral phenotype with a neural implementation [20]), might have had an influence in the establishment of this take on vocal learning. The existence of a direct connection from forebrain to phonatory muscles, allowing for fine control of those structures, is indeed an appealing idea on which to build.

There is, however, empirical evidence of vocal learning abilities in other species outside of the canonical list. Such evidence is usually behavioral and not an attempt to show direct forebrain control of phonatory muscles. Indeed, there is work questioning that such a connection is a necessary condition for vocal learning in the first place [21, 22].

The vocal learning continuum and beyond

In light of this, we think it worth discussing the “vocal learning continuum hypothesis” (VLC) [11], which categorizes species along a continuum of increasing vocal learning complexity. This is a valuable idea that goes against the traditional dichotomous view of vocal learning, according to which a species is either definitely a vocal learner or not at all. However, it too relies on production of novel calls through imitation, subserved by forebrain control of phonatory muscles, to determine the distribution of vocal learning abilities across species. Such an approach is therefore not representative of the diversity of vocal learning behavior across the animal kingdom [23, 24]. This diversity pertains not only to species for which there is recent evidence of vocal learning but also to the “well-established” vocal learning species, namely birds [25].

Indeed, we find that the way species learn to produce their communication signals should form the set of criteria that makes a species a vocal learner. The exact nature of the vocalizations and the neurobiology are of course extremely important, and they will allow for much more precise evolutionary work, but one must not lose sight of the fact that vocal learning is a behavioral phenotype, with learning as the most striking aspect.

Given this state of affairs, two ways offer themselves for future studies on vocal learning: (1) perpetuating the bifurcation between canonical and “negligible” vocal learning species or (2) turning attention to the behaviors observed and assessing them in the context of a broad sense of vocal learning, as opposed to dismissing them on neurobiological grounds alone.

Recently proposed frameworks and reflections also show some concern with this question and call for more wide-ranging perspectives on vocal learning (e.g., [26–28]).

In the remainder of the present paper, we start by going over the VLC and point out some limitations. We then point to evidence from “noncanonical” species that reinforces these limitations and conclude by outlining an extension to the VLC highlighting approaches to vocal learning that can help overcome them.

Bidimensionality

Because vocal learning categories are determined by the existence and strength of a particular brain circuit in the VLC, this makes it a bidimensional system (see [28] for some recent discussion on the same issue).

The particular brain circuit is, to put it in simple terms, a direct connection from the forebrain to phonatory muscles, and it is thought to be present in some form in canonical vocal learning species. The Kuypers/Jürgens (KJ) hypothesis [8, 32] posits that such a connection is necessary for the kind of motor control that is required for vocal learning, and the VLC tacitly relies on it. This idea had already been made popular before (e.g., [33], among others), but it was perhaps made more widespread in the work of Jarvis (e.g., [11, 34]), and Fitch (e.g., [8, 35]), who named the hypothesis after two scientists who made important contributions to primate neurobiology [36–38]. In other words, even though the VLC is a more nuanced conception of how to ascribe vocal learning across species because it allows intermediate steps, it is still limited in the sense that it has the forebrain-to-phonatory-muscles connection as the sole predictor and allows for variation only in that dimension. It is relevant here to recognize the role of a direct forebrain-to-phonatory-muscles connection as a necessary ingredient in the VLC; proponents of the VLC are of course well aware that it alone cannot explain away vocal learning as a whole. Other abilities and traits are involved, such as auditory learning [11], but the VLC is not concerned with them.

An analogy to the bidimensional nature of the VLC would be a slider in a physical machine or a computer program that controls a parameter, and by sliding it back and forth, the output is changed. In this case, the parameter would be the strength of a forebrain-to-phonatory-muscles connection, and the output would be “less” or “more” vocal learning. If the slider is at position 0, we get no vocal learning. If the slider is at the maximum value, we get “high-end” vocal learning.

There are two ways in which this bidimensionality is problematic. The first is that it leaves out capacities and constraints at other levels of analysis [18, 24, 28], which might or might not go hand in hand with this brain circuit. This is well captured by the following questions, taken from [28]: (1) What makes a species a vocal learner? (2) When is vocal learning employed? (3) How can vocal learning be expressed by the organism? (4) Who (else) is capable of vocal learning? And (5) why did vocal learning evolve?

Recent empirical evidence shows that species that do not or are thought to not have relevant forebrain control of phonatory muscles can be vocal learners. This brings to the fore other ways in which species may achieve what is in effect vocal learning behavior. Testing of species whose vocal learning capacities are unknown or supposed not to exist still yields surprises.

The other problem is that if a goal (or even the main goal) of comparative work is to derive information about the evolution of traits and species, we cannot expect a single aspect (e.g., a single genetic change or a single brain connection) to offer a realistic picture of how the trait evolved [39]. Even if the empirical evidence established that only vocal learning species in any one sense consistently have a certain brain connection and vice versa, evolutionary aspirations would still require a more complex explanation. In the realm of complex traits, there is always a cascade of effects with far-reaching implications [40]. It is also the case that even homologous behaviors don’t necessarily share a neural mechanism: there can be genetic changes affecting circuitry with no change in behavior [41].

More on brain wiring

The discovery of particular wiring (see Box 2) made it possible to attempt a principled, brain-based separation of strictly innate calls in a way that’s shared among all mammals tested from calls that are controlled volitionally. However, even in the very strict sense of learning of novel vocalizations through imitation, it is not known beyond doubt that this is a necessary condition. For example, there are reports of learned, voiced calls in the orangutan (Pongo spp., a species that supposedly lacks the relevant connection [42]). It is also not clear whether the connection is sufficient (within reason) either; mice (M. musculus) apparently have the circuit but do not produce novel sounds through imitation [29, 30], and perhaps more interestingly, recent work shows that female zebra finches (Taeniopygia guttata), which do not produce learned song, have “male-like” song pathways [43], so the narrative is not totally compelling. Furthermore, there is work showing the involvement of other structures and pathways in the learning of vocal behavior in a relevant manner, such as the cerebellum [44], the periaqueductal gray (PAG) [2], or the ventral tegmental area (VTA) [45]. It is also not entirely clear why vocal learning, a phenotype whose most interesting aspect is arguably the learning part, must be limited to a certain kind of vocalization, namely the kind that requires fine control of the phonatory muscles (what is usually referred to as “phonation”). Moreover, it is important to bear in mind that any one connection does not exist in isolation; each brain region involved will be part of several other connections, each with its own complex evolutionary history.

It thus seems that, although phonatory muscle control is obviously a very useful ability, relying on the KJ hypothesis alone might not give us a good indication of the basis of vocal learning and how widespread the phenotype is.

Complexity considerations

The VLC also aims to represent various degrees of vocal learning complexity. But as we will see on at least three counts, it does not do so satisfactorily.

Bengalese finches (Lonchura striata domestica) are the domesticated strain of the white-rumped munia (L. striata). Domestication has been claimed to increase vocal learning complexity: if a “wild” species is already a good vocal learner, it becomes a more complex vocal learner after undergoing domestication [52]. In the case of the Bengalese finch, for example, this happens despite the fact that this bird species was not bred for its song. It is possible that imitation—crucial per the KJ hypothesis and, concomitantly, the VLC—could actually be detrimental to syntactic complexity. Compared with their wild counterparts, Bengalese finches display higher unpredictability and syntactic complexity in their song because not only do they imitate their tutors (partially) but they also improvise, resulting in what is, in effect, lower imitation fidelity [53].

The full “classic” circuit of vocal learning involves a posterior pathway for vocalization and an anterior pathway for learning [34]. Besides those pathways, parrots were discovered to have a “shell” song system in addition to the “core” song system in all vocal learning birds [19]. A larger shell system relative to the core system is associated with parrot species that have more “complex” vocal learning abilities, and vice versa. Moreover, this shell system has mostly intercortical connections, as opposed to the direct connection to the motor neurons characteristic of the core system. It seems, then, that parrot species with a larger shell system have an edge in the VLC, but this edge is not related to the direct connection the VLC rests on. This poses a challenge to the VLC as it stands because it requires the addition of an extra factor (say, adjacent “song/speech” nuclei with intercortical connections or even just “strong intercortical connections”), opening way to a much less restrictive VLC because more factors would be added as needed for capturing differences between species, departing from the bidimensionality we have already alluded to. Indeed, Chakraborty and Jarvis [54] acknowledge it might not be straightforward to reconcile the core/shell system with the VLC.

Finally, in the VLC, humans alone are considered high-end vocal learners, whereas parrots are classified at a level just below, referred to as complex vocal learners. This is purportedly because of the higher syntactic complexity in human vocalizations, but this does not rest on the criteria for categorizing species along the VLC (presence and strength of direct connection to the phonatory organ and imitation). Language complexity need not even be instantiated in vocal behavior; it is well established that the linguistic capacity is the same in sign language (see [55]). It could be that, indeed, humans are the most advanced vocal learner, but this is not possible to discern from the criteria used in the VLC. It might have more to do with the process of cultural transmission and not with anything “vocal” [56]. In a manner similar to Bengalese finches, it has been hypothesized that the increased prosociality that characterizes domestication allows for the jump in complexity to take place (see [56] for discussion).

Imitation and de novo vocalizations are not the whole story

An important step, in our view, is to adopt a view of vocal learning behavior not necessarily focused on imitation.

The production of de novo vocalizations (new in a species repertoire) and, furthermore, doing so through imitation is often taken as the golden standard when assessing vocal learning abilities. This is problematic because imitation is one possible means of displaying vocal learning behavior. Indeed, diverging from imitation is also a common phenomenon in vocal and cultural development [23].

Perhaps a more productive conception of vocal learning is looking at learned vocal behavior as having to be acquired in some manner over developmental time, especially in contingent ways (that is, dependent on experience and not a “certainty” given the initial state of the organism).

There are interesting cases that illustrate vocal development by diverging from the tutor song; that is, by the countering of or lack of imitation. Infant marmosets, for example, develop vocal learning abilities through social reinforcement from parents, not imitation. This leads to more control of the vocal apparatus, which allows them to produce lower entropy calls [24, 57].

Canaries (Serinus canaria domestica) trained on atypical song imitate it at first but, when reaching maturity, shape it into the species-specific song they were never exposed to [58]. Another well-known example is the de novo emergence of zebra finch song not by imitation but, instead, by the approximation of wild-type song over a couple of generations by birds reared in isolation, with no exposure to singing tutors [59].

Evidence of this kind is good indication that vocal learning is not driven (solely) by imitation and that vocal learning ability is characterized also by behaviors that suppress imitation.

Evidence from species outside of the canonical vocal learners list

Opening up to more-permissive definitions of vocal learning goes hand in hand with opening up to the study of more species. A decent amount of evidence for vocal learning outside of the canonical list has been put forward, especially in recent years, with primates as the most representative of this trend, and some work on rodents. They moreover deserve special attention because there is resistance to taking this kind of evidence into account. Other species are more quickly accepted, perhaps because they employ imitation, and neurobiological information on these species is given a great deal of importance, given its scarcity. A good example of this is the African elephant (Loxodonta africana), which quickly entered the accepted list of vocal learning species [16].

The logistic difficulties in keeping and studying larger species as opposed to birds and other (usually smaller) species might also bias positions against, say, primate vocal abilities, leading to a situation in which absence of evidence might be mistaken for evidence of absence. Rodents present challenges on their own (e.g., several species produce ultrasonic vocalizations, which pose further challenges, and there is a lot of interindividual variation [60]). Studies on birds have unsurprisingly dominated vocal learning research (see data in [28]). Great ape language acquisition projects (e.g., [61]) might also have contributed to this state of affairs, given their varying goals and approaches, which usually had to do with finding some form of human language, as well as difficulty in interpreting their results. In Box 3, we summarize some evidence that we think deserves, at the very least, attention if vocal learning is to be understood as a phenotype that’s characterized by the learned modification of calls, with the exact nature of the calls being an important but not (dis)qualifying feature. A more complete (yet not exhaustive) list can be checked in S1_Text.

We believe that evidence of the kind we review here has only been neglected because of its nonconformity with the KJ hypothesis. We find that one way of getting a full picture of vocal learning is placing the focus on observing the behavior, without preconceptions of what should allow it, and then proceeding with the mechanisms. As put by Krakauer and colleagues [20]: “The neural basis of behavior cannot be properly characterized without first allowing for independent detailed study of the behavior itself.”

The relationship between a neural structure and a behavior is not one of explanation of the behavior [20]. This is the case even if that relationship is consistent, which in the case of vocal learning and according to evidence we reviewed, it might not be. Although some neuronal implementation will of course be in place, there is no indication that the vocal learning phenotype can be equated with a particular one given that there is not even a consensus on what the behavior encompasses. In the face of paradoxical evidence (e.g., primates displaying vocal behavior they are not “supposed to” have), rejecting the behavioral evidence instead of revising the neural hypothesis will not lead to understanding of the trait. It is in this sense that we think it is important to have behavior as the entry point to the study of vocal learning. If the VLC is extended beyond the specific neural substrate that is taken to allow the direct control of the vocal organ, we could gain a better understanding of the phenotype.