Development and evolution of song in swamp sparrows

Following an introductory section, this page has four parts: (click on a number if you wish to jump forward to a section)

1. our current understanding of vocal learning, particularly that which has developed from work on sparrows

2. vocal learning experiments on swamp sparrows in collaboration with Jeff Podos

3. functional tests of swamp sparrow song in collaboration with Dana Moseley

4. how our developmental studies might be developed in order to address important questions about human language


I spent five years with Bob Payne at the University of Michigan, who had studied songs of indigo buntings and African finches for three decades, and I somehow managed not to get into vocal communication. Birdsong in grad school was like U2 in high school-- too many people going ga-ga over it, I'm not falling for that. But eventually my defenses (against both) fell. I was mainly interested, as I still am, in the factors that influence trait evolution, and the fact that oscine bird song was learned prevented me from seeing it as a good way to study evolution. I can see now that three things helped me out of this typological rut.

First, Richard D. Alexander, who was a valuable mentor to me throughout graduate school and since, continually reminded me that learned traits are no less products of evolution than traits whose development is (presumably) guided more rigidly, that learning itself is a trait that has a genetic basis and evolves, and that learning tends to be adaptive and thus is usually if not always biased towards certain objects and away from others. Second, while looking for independent postdoctoral opportunities I discovered Jeff Podos and his Animal Behaviour papers that demonstrated, among other things, a major syntactical abnormality ("broken syntax") in the song of some individual swamp sparrows that were trained as youngsters to sing songs faster than any swamp sparrow could actually sing. By pushing young birds against a performance limit, Jeff elicited new song structure. Third, around the same time I read a bit of news in Science including a comment made at a bird song meeting at Rockefeller University... I think it may have been by Emilia Martins... that the most important unanswered question in the field is how species differences in bird song arise. It had always seemed to me that songs of different species of Melospiza sparrows, for instance, were so hugely different that addressing the question of how and why they diverged was impossible.

But Jeff's work made me wonder: what if females with their choosiness about male song have pushed males against limits of their abilities in the same way that Jeff's experiments did? And what if any strange songs that males might produce as a result were (very occasionally) actually liked by some females in some corner of their range? They might mate with those strangely singing males, the subsequent generation of males may sing like their fathers, the daughters might tend to like the songs of their fathers just like their mothers did, and voila, we have the potential for divergence in the structure of bird song. Regardless of the particular mechanism (a few others have also been proposed), as the last sentence in Jeff's first swamp sparrow paper says, performance limits leading to song abnormalities might "provide a mechanistic basis for an evolutionary novelty".

In this way I finally saw the reason why an evolutionary biologist should want to study the way birds learn their songs: by examining the structure of development we can gather clues about evolution. Rather than learning being an impediment to this project, the very objects and limitations inherent in an organism's learning process arose through evolution. If we understand them we might wander into interesting uncharted territory with regard to the evolution of learning and learned behavior. And since we understand the development of bird song probably better than any other complex learned behavior on earth, there seems scarcely a better model system.


How birds learn their songs: what we know

Once upon a time there were two competing models of how animals (including humans) learn how to speak or sing. One (after B. F. Skinner) proposed that just about any sound can be learned as long as there is adequate conditioning or instruction. The opposing view (after Noam Chomsky) hypothesized an "innate core" of communication elements, which is whittled over development such that useful elements are retained and unused ones are eventually discarded. The empirical work on bird song and human speech doesn't give much support to either of these extremes, however. Peter Marler and others developed an alternative known as the "active vocal learning" model, which is a sort of middle way in which both instruction and selection are important. In a perceptual phase of learning an organism needs to be instructed (i.e., to hear) what to sing or say. In a practice or sensorimotor phase, selective retention of the more useful elements occurs. In humans these two phases overlap-- babies memorize sounds they hear at the same time they are practicing them. In many songbirds, however (including the sparrows on which Marler, his student Nowicki, and Nowicki's student Podos work) the phases are separate in time: the birds learn by instruction in their first summer, but they practice and select the most relevant sounds the following spring, their first breeding year. Another important feature of sparrow song learning is that the window of opportunity closes in early life: after a few months of age, sparrows memorize no new songs; and after a month or two of practice, they "crystallize" or settle down on a final form their own songs will take for the rest of their lives.

The active learning model of vocal development:

Active vocal learning has solved a lot of inconsistencies and seems to be largely correct, but recent developmental work has suggested that we need an additional modification. The active vocal learning model assumes that the neural representations (memories) of the target sounds don't change after an animal learns them in the perceptual phase. The sensorimotor phase is then all about trying to match those memorized sounds as closely as possible. Jeff's swamp sparrow research challenges this assumption. He presented young male swamp sparrows in the lab with high trill rate songs for them to learn, and the birds learned the songs but were unable to reproduce them at their rapid rates. Several lines of evidence indicated that these birds ran up against motor limits, meaning that they were physically unable to produce the songs at that high speed. Instead, they introduced modifications to the fine structure of the song, such as reduced trill rates, note omissions, and a novel “broken” syntax where a song is divided into sections with pauses in between. By following the development of these songs (fortunately, birds practice out loud), Jeff was able to infer that the songs were memorized at their original fast trill rates, and were produced that way (albeit imperfectly) in the beginning of development. Eventually, however, the birds introduced the modifications that made the songs easier to produce, and future practice narrowed the output to that version, instead of to the originally memorized version. These modifications were beyond those that animals normally make as their neural and motor systems mature. Jeff calls this new mechanism the "calibration" model because the birds seem to be adjusting the goal of their practice according to their particular performance abilities. This finding suggests that the memory or "template" that birds form of the songs they hear as fledglings, and which becomes the goal for later practice, is not carved in stone. This template is apparently somewhat malleable and can be adjusted while birds are practicing. What this finding does to the active learning model is to add a feedback loop between the sensorimotor phase and the memory of songs heard in the perceptual phase.

The revised active learning model:

Again, this feedback loop is not the same thing as birds listening to their own song, comparing it to their memorized version, and modifying their output to correct inconsistencies between what they sing and what they had learned. Birds do this also, of course, but this feedback loop is different-- it's not changing how you aim, it's changing the target. Jeff's results suggest that when birds learn a song and later find they can't reproduce it properly, they don't just struggle for the rest of their lives to get as close as they can, nor do they just give up and produce shabby song; rather, they change what they're aiming for to something they can produce consistently, and crystallize that.

How to perform an advanced vocal learning experiment in three uneasy steps: Take swamp sparrow nestlings from the wild into the lab and rear them by hand... ...Train them every day for over a month on swamp sparrow songs that have been digitally manipulated in order to vary certain features of interest... ...Then, next spring, place them into sound isolation chambers and record the songs they sing!


How birds learn their songs: what we're testing now

My recent vocal learning experiments with Jeff Podos were designed to further develop and test this revised active learning model. For instance, there is good evidence that the combination of rapid trill rates and a large vocal range (frequency bandwidth) in a song is difficult for males to produce, and is preferred by females. These two effects make sense together-- if females are interested in males of high quality, they might be able to assess this by listening to aspects of song that are challenging to males. Anyway, we know now that the calibration feedback loop in the revised model is used when males are faced with difficult songs. Will they modify their vocal target in other situations as well? Our experiments examine three particular possibilities:

1. What if males are presented with songs that are not as difficult as they could produce-- do they optimize them to impress the females, or do they imitate the songs strictly as they hear them? Sexual selection theory would predict that males would attempt to maximize their reproductive success by departing from strict imitation if they can improve on the songs they have learned. We are testing this possibility by training a group of males with songs whose trill rates have been modified to be slower than normal.

2. What if males hear songs that are species-atypical in some way? Individual wild birds, including swamp sparrows, often produce songs with odd features such as irregular trill rates, changes in frequency range over the course of the song, or changes in amplitude (volume). Do males that learn from these songs "correct" these irregularities when they are young, or do they just imitate the models the way they are? The fact that most swamp sparrows sing very typically would suggest that they do not blindly imitate but that they have a hitherto unknown error-correction mechanism built into the developmental system. We are testing this possibility by training a group of males with songs with artificially accelerating or decelerating trill rates.

3. We would also like to know about learning biases. If males are given some high performance songs and some low performance songs during early development, do they choose to memorize the higher performance songs during the perceptual phase, or do they simply use their sensorimotor phase to optimize whichever song they happen to memorize? Do swamp sparrows have any other biases regarding note type, trill rate, number of notes in a trill, or other structural features of song? Our field has concentrated so much on the learned aspects of song that we have mostly ignored the fact that many aspects of development, even of vocal development in songbirds (and humans), are not learned. Any unlearned factor that influences development of song can be considered a learning bias. We are testing this possibility by training a group of males with some slowed songs and some sped songs.

All experimental groups are also presented with two songs at natural trill rates.

Between 2005-2008 we collected the data for these questions, and now we are analyzing the data and writing at least three papers. Our first manuscript will soon be published, and documents the tendency of young swamp sparrows to learn a variety of songs but to depart from the models to optimize performance.  Our second manuscript based on these results will more fully explore the ways in which inherited learning biases influence the development of bird song. Our third manuscript will be based on a more thorough analysis of the developmental stages these birds traversed as they were practicing. 

One exciting possibility for the future is to compare the songs of birds we train in the lab with those of their genetic fathers in the field, whose song they have never heard. Is performance, rhythmic consistency, or any other singing feature inherited? These experiments thus provide us with an opportunity to compare the influence of nature vs. nurture on vocal development and song.


What are the songs saying in nature?

To complement the laboratory vocal learning studies, I have been collaborating with University of Massachusetts graduate student Dana Moseley on field studies for the past five years. Our purpose is to decode the songs, i.e., to find out what certain song features actually mean to swamp sparrows. We do this primarily by experimental tests of how differences in these song features (especially trill rate, but we may use similar methods for frequency bandwidth and rhythmic consistency) translate into differences in behavior of males and females when they hear these songs. In this way we can approach the biological function or communicative value of those song features... essentially translating bird song into English, at least in a functional sense. 

There are three aspects to this project:

1. Assess song variation by extensive recording in the wild, especially noting the song features of genetic and social parents of both the laboratory-reared birds and of other wild birds in the population.

2. Experimentally test the role of trill rate by playing the same song to a male on his territory on two different occasions, one at normal trill rate and one at a sped trill rate. This simulates a territorial intrusion by another male, and we would expect the resident male to respond according to the level of threat that the intruding male song represents. We have developed a series of predictions of male response that take into account both the subject male's own performance relative to that of the model song, and the threat that this difference would indicate. We have now completed three seasons of such playback experiments spearheaded by Dana, and she is currently analyzing the data.

3. Raise females in the laboratory and train them on modified songs in the same way as we have the males. Then test for their mating preferences by first conditioning them to relate songs to mating decisions, and second implanting them with the sex hormone estradiol and playing songs to them when they are in breeding condition and observing their behavior. This year Dana has been able, for the first time among laboratory-reared sparrows, to elicit stereotyped sexual behaviors in females, most extremely the back-bending copulation solicitation display posture. In this year's study Dana played these birds some songs that they were trained on when young, and some novel songs, to test for a familiarity or training effect on preference. The songs also have artificially modified trill rates, allowing us to test whether the females express an unlearned preference for faster songs. This project is essentially Dana's, and she is currently (with undergraduate assistance) working through hours of video recording to determine what song features elicit the greatest sexual responsiveness.

We have performed the field portion of these studies in a population of swamp sparrows on Prescott Peninsula, Quabbin Reservoir, Massachusetts, with help from undergraduate assistants, other graduate students, and our friend and postdoc/lecturer/environmental consultant Steve Johnson. For more information about this field site see the About Me page.


An application: bird song and human speech

A number of striking parallels have been uncovered between the development of bird songs and human speech. Both babies and young birds copy sounds from acoustic models, and those sounds are stored as neural representations before they can be produced. Learning in both systems occurs most readily during a limited temporal window of opportunity, corresponding to patterns of neural growth and development. Recall of those sound memories in both songbirds and humans occurs only gradually, as animals first produce poor copies of models (“babble” in humans, “subsong” in birds) in early development. Most relevant to our research, auditory feedback in both systems enables birds or humans to refine their vocal output, through a process in which vocal output is repeatedly heard, compared to stored memories of song models, and refined to achieve increasingly accurate copies. Surprisingly, despite the millions of years of distance between these two lineages of complex vocal learning, both genetic and neurobiological studies have uncovered similar control pathways for vocal development in songbirds and humans. Because of these parallels, and because of ethical restrictions on human experimentation, studies on birds have provided some of our most valuable empirical insights into speech development (see my publication list for articles that develop this topic further).

Although results from the swamp sparrow may not be directly applicable to other species, the use of a songbird species as a model system can inform new hypotheses and suggest fruitful avenues of research relevant to other songbirds as well as humans. For instance, in order to classify and treat human speech disorders, we need to understand the development of vocal communication in early life. A recent review concludes that “acquisition research lies at the heart of the modern study of language” (Kager et al. Constraints in Phonological Acquisition, 2004). Specifically, researchers of human speech development have recently encouraged as an immediate goal the clarification of the distinction between factors that guide and limit speech development at the perceptual phase (comprehension-specific constraints) and those that act at the sensorimotor phase (production-specific constraints). Improving our model of vocal learning, and especially clarifying the role of the sensorimotor phase, is a precondition for addressing some persistent scientific and clinical issues. One of these relates to the recent neurological discovery that the perceptual and motor circuits are interlaced in the brains of both songbirds and humans. Support for a feedback loop between production and memory as a normal feature of vocal learning, as the revised active learning model proposes, could help explain why such close connections exist between these two systems. Support for this feedback loop would also suggest directions for future neurophysiological investigation, especially regarding the search for the "template" or song memory in birds. For instance, if the revised model is supported, the template would be best viewed not as a location of passive storage, but as an entity that is connected dynamically with the pathways that underlie sensorimotor learning.

The kinds of experiments we perform on swamp sparrows will also shed new light on three lingering general questions about the functions of the perceptual and sensorimotor phases of development.

1. What is the relative importance of imitation versus active modification of memorized models during vocal development? Studies of atypical speech development in humans have shown that many individuals alter model speech to suit their own capacities. Experiments could help determine the relative importance of imitation and improvisation, but such manipulations are not appropriate for humans. Our research, on the other hand, is directly relevant to this question. Although not all the data have been analyzed, it seems that improvisation to match production ability is the norm with swamp sparrows, and this this is suggested by human observational research as well.

2. In the event that birds and humans do modify their vocalizations because of constraints or to optimize from their models, at what point in vocal development do these constraints or biases occur? The most straightforward division is between the perceptual and sensorimotor phases; applying the issue of constraints to this division addresses serious questions in human speech pathology. For instance, children often apparently learn units of language that are never incorporated into their communicative repertoire; in these cases, as one clinician has recently asked “does perception or production reflect the state of phonological competence in the child's emerging linguistic system?” (Joe Pater, in the Kager edited volume cited above). Do the children fail to remember the language unit, or alternatively are they unable to utilize it? An answer to this question would help determine whether and how treatment of a speech pathology can proceed, but in many cases there is an insufficient basis to assign the constraint to either phase of development. This may be partly due to the difficulty in distinguishing the precise roles of the two phases. Our research directly addresses this issue in birds, where the development of competence in vocal communication can be monitored more thoroughly, and where several sources of production constraint have been studied.

3. How strong of a role does social learning play, relative to other factors, in the development of the final form of speech? In addition to this question’s evident importance to psychologists and society in general, the position taken by researchers on this issue can strongly influence the level and nature of clinicians’ (and presumably educators’, social workers’, and parents’) concern for children’s speech development. This question is often posed in terms of learned vs. innate aspects of language acquisition. Recent reviews have expressed significant interest in research that would clarify the ways in which social learning and unlearned strategies cooperate in guiding the developmental process, whether in birds or in humans. By controlling and altering the environment of young birds, our research allows us to measure effects of social learning independent of other developmental influences that do not vary with socialization. Any difference in developing or crystallized songs between birds exposed to control and modified song models represents the influence of social learning. On the other hand, features of song that are robust to our experimental design are either unlearned species-specific features, or involve learning from cues that are consistent across treatments.