Apes Don't Imitate; People Do

Rich Little is a master mimic, a peculiarly human talent.

One of the things that distinguish ape cries from speech sounds is that ape cries are holistic; that is to say, the whole sound is unique, so that a pant is unmistakably not a scream, and a hoot is still a third noise. Meanwhile all languages build on a particular set of sounds that are used over and over: e.g., remember, member, membrane, reign, terrain, terror, tear up … Somehow our ancestors seem to have shifted from making holistic sounds to making combinatorial ones. How did they manage that? (Sign languages raise the same issue; ape gestures are holistic, as are those of hearing humans, while sign languages build on combinations of gestures.)

Conscious invention might be a possibility, except that anybody who has worked with linguistically naïve people (e.g., young children and illiterates) knows that they commonly do not realize that their words are combinations of a small set of sounds. Thus, even if a genius did come up with the idea of combining sounds, there is still the problem of how the system spread throughout the population of non-geniuses and how ordinary children pick up the combinatorial system without noticing how the system works. I have always found it difficult to think clearly about this somewhat technical subject, but thanks to a couple of recent papers my thinking is (I hope) a bit more focused.

Bart de Boer and Willem Zuidema have recently co-authored two papers on the evolution of combinatorial phonology in the Journal of Phonetics  and Adaptive Behavior. They have worked out a computer simulation to test their ideas. Simulations are not perfect, but they can be very useful in thinking about what is required to evolve some novel function.

Their simulation, in essence, produced a signal and determined which signals were easiest to learn and reproduce. Their computer created a series of “agents” with a repertoire of randomly defined signals. Simplified a bit for this blog, the simulation, or game, has several steps:

1. One agent generates a signal in its repertoire, with some random variation in its form.
2. Another agent searches its own repertoire to find the closest match to the original signal. The agent then generates that signal with some random variation added.
3. The original agent takes the imitation signal and sees if it best matches the original signal, or is more like some other signal in the original agent’s repertoire. If the imitation signal is judged to be most like the original signal, the imitation is judged a success. Otherwise, it is judged a failure.
4. In the case of a success, each agent replaces the original signal in the repertoire with the one produced. Otherwise, the signal is unchanged.

An example of a round in the game: (1) agent A randomly selects bakshish from its repertoire which is randomly altered to bikish; (2) agent B searches its repertoire and and settles on lickerish which comes out as lackish; (3) Agent A checks and can find no ideal match but accepts lackish as closest to bakshish; (4) Agent A replaces backshish with its signal bikish while Agent B replaces replaces lickierish with lackish.

This may seem like a trek to nowhere, but both –shish and –erish have been simplified to one –ish. Slowly, without any agent intending it, the signals are taking on a series of shared sounds and recognition comes to depend more and more on matched combinations. Over many rounds, a holistic set of signals becomes combinatorial.

So, you may be wondering, if it is so automatic, why are most animal signals holistic rather than combinatorial? The answer seems to be that very few animals engage in any variation of the game studied by the authors. Most animals do not learn their signals through imitation, nor are they trying to duplicate the sounds of their neighbors. In cases of imitation-based signals, as with songbirds and whales, combinatorial systems are used.

One important difference between human behavior and the imitation game in the paper is that producers judge their own sounds rather than the sounds of others. I’m basing this position on a description of how zebra finches learn to make their song. (See: Birds also use FoxP2) The FoxP2 gene contributes to their brain’s ability to compare the sound they hear with the sound they produce, so that after many practice tries they can produce the sound they hear quite accurately. This approach makes more sense psychologically since social learners typically work to be like their teachers and do not judge them.

Either way, the process seems to imply a series of stages. First, is holistic production, but a production in which the form does not much matter. If it did matter, there would be selection pressures to keep it as it is. This might seem like tall order, but we see it in infants who babble at about seven months of age. They make their sounds and it does not matter what sounds they make. Later, they start playing their version of the imitation game and learn to make combinations.

I have always thought that babbling was the critical shift toward phonology, but now I see there was a second equally important one: wanting to be like your teacher. This step is just assumed by the simulation, but at least we have identified a pivotal point. Along with a need to make sounds, cooperate and trust, there has to be a drive/urge/instinct/whatever-you-call-it to imitate the signals of another. Where did that come from?