The fidelity of iterated vocal imitation
Pierce Edmiston, Marcus Perlman, and Gary Lupyan
EvoLang 11
Ancestral charades
Probably the easiest thing to do is to imitate them. You could act out their movements, or easier still you could just imitate the sound they make.
It turns out that people are pretty good at vocal imitation. So let's see if you can tell what this person is thinking about.
Can words emerge from imitations?
- quack, sizzle, ooze, hiss, boom
- But can we watch it happen?
Can you start with an imitation and gradually form something that is more wordlike simply by having people repeat the imitation over and over?
We have some suggestive examples of words with origins in imitations in English, like onomatopoeia, words like quack, sizzle, ooze, hiss, boom.
But can we watch it happen in real time, and gradually, as a function of iteration?
The experiment was very simple: participants clicked on the top icon to hear a message, and then they clicked on the bottom icon to make their recording. They were instructed to repeat whatever they heard as accurately as possible.
We didn't want to use animal sounds or anything that had a lexical form, so instead we used sound effects, like the tearing of paper, the clinking of glass bottles together, water sounds.
Here are two examples from an early version of the experiment.
We picked four categories of sounds: glass sounds, tearing sounds, water sounds, and zipper sounds.
Within each of those categories, we had four distinct seeds, so we had 16 seeds total, but any given participant only made four responses: one in each of the four categories. So the green circles here represent the first subject in the experiment.
Measuring imitation fidelity
To do that we had a different group of subjects listen to each of the imitations, and see if they can guess which seed lead to that imitation.
I'll call this the “Guess the seed” game, and in all cases people were given four choices, and they had to pick the one that most resembled the imitation that they were listening to.
In this case, the highlighted imitation came from the “c” seed, so the correct answer to the Guess the seed question is “c”.
If people are above chance at guessing the seed, we can say that the imitation has some fidelity to the original source. However, if you played the game of telephone as a kid, you know that the message that comes out the end often barely resembles what you started with. If people are guessing at chance, we would say that the imitation lost fidelity.
So that's what I mean by imitation fidelity: the ability to relate a particular imitation back to the original source.
And of course they are totally blind to the design of the experiment, what generation the imitations came from, and they don't get any feedback on their performance.
Are imitations becoming more wordlike?
One property of words that I think about a lot is that words are categorical, whereas imitations are more specific. If you are imitating a dog, you are imitating a particular dog. An imitation of a yorkie sounds different than an imitation of a German shepherd. But the word dog is more general, it leaves the type of dog unspecified.
This has a specific prediction for what should be happening with the imitations. If they are becoming more wordlike, the imitations should lose individuating information more rapidly than category information.
Remember that in the design of the experiment we have four categories of sounds, and four seeds within each category. Category information is the ability to distinguish an imitation from any of the seeds in a particular category from the other categories.
Individuating information is the ability to identify the exact seed that lead to the imitation within a particular category.
Now, I designed the experiment in this way so that I could take each imitation and have it rated in different version of the guess the seed game. And we can use these ratings to address the question of whether the words are becoming more wordlike.
So now I'll show you how I structured the Guess the seed games.
You should recognize the design of the telephone game with the four categories and the four seeds within each category. And remember for the Guess the seed game people are given a single imitation – here the imitation is orange, and they have to pick which seed the imitation game from.
So these three Guess the seed games differ only in the choices that the participant is given.
Let's start with the easiest condition at the top. In this condition the correct answer is the actual seed that lead to the imitation, and the distractor choices are from different categories. I call this the “Category match (true seed)” condition.
The second condition, the “Category match” condition, is a bit harder. You're still given seeds from four different categories, but the correct answer is not the “true seed” but a different seed from the same category.
Performance in this “Category match” condition is really important. If an imitation leads to above chance performance in this condition, it means it's a good cue for the category of sounds as a whole.
The last condition is the hardest. In the “Specific match” condition you have to match the true seed, but the distractors are from the same category. Now, an imitation that is a good cue for the category as a whole will not perform well in this condition, because if it is a good cue for all of the seeds, then you probably can't tell which exact seed lead to that imitation.
Ok, so again: we collected all of those imitation in the telephone game and then for each imitation, we had it rated in each of these three conditions. And now we can see how good the imitations were.
The first thing to notice is that performance in all three conditions is above chance, which requires both good imitations and good guesses.
The next thing to note is that the order is what we would expect. The category match where you are given the true seed is the easiest, and the specific match is the hardest.
Now this is the first generation, and I want you to think about what is going to happen when we look across generations.
Again, the prediction is that the imitations will lose the individuating information more rapidly than they lose the category information. This plays itself out in two ways.
First, focus on the two category match conditions. We can call the difference between them the “true seed advantage”: it's the benefit you get as a guesser when one of the category options is the actual seed that generated the imitation. This true seed advantage depends on individuating information; it depends on the imitation in some way resembling a particular seed more than the other members of the same category. So if the imitations lose the individuating information, we would expect this true seed advantage to decrease.
Ok, now focus on the difference between the category match condition and the specific match condition. The specific match condition obviously depends on individuating information, and the category match condition doesn't, because the correct answer is not the specific seed that lead to the imitation. So here we expect the difference between the two conditions to get bigger: performance in the specific match condition should drop more rapidly than performance in the category match condition.
Are imitations becoming more wordlike?
But are they becoming actual words? Can you write them down?
Transcriptions
Conclusions
Early stages
Click bait
Thanks!
