If you hear a language you do not speak, it is hard to work out where one word ends and the next one begins. So finding the boundaries between words in continuous speech is a pre-requisite for acquiring your native language and for learning another language. A recent experiment found, for the first time, that domestic dogs can distinguish boundaries between words in human speech in a way at least partly comparable to humans.
Frequencies and boundaries
A good way to start identifying word boundaries is to learn which syllables are more likely to follow each other. From as early as 8 months, human infants can use such regularities spontaneously to learn to spot where words begin and end—long before they learn what those words mean. To do this, infants make complex calculations to keep track of syllable patterning: syllables that often appear together are probably words, and syllables that do not often appear together probably aren’t words.
For example, consider the phrase scary dragon, made up of a sequence of 4 syllables: sca-ry-dra-gon; the syllables forming a single word, like sca-ry in scary, are more likely to occur next to each other than syllables forming part of different words, like ry-dra.
Can dogs use these frequencies?
Dogs live among us and hear us talking all the time. Can they also use statistical regularities in speech to spot the boundaries of words? To find out, Hungarian researchers at the Department of Ethology at Eötvös Loránd University measured dogs’ brain activity. Dogs listened to a continuous speech stream in which some syllables occurred together more often, like parts of a single word. EEG measurements showed that dogs’ brains distinguish sequences of syllables forming these “words” from random sequences of syllables.
How do dogs do that?
As shown by previous research, humans use at least two types of distributional statistics in segmenting a continuous stream of speech into words:
- co-occurrence frequencies (how often words occur together). Humans can already compute these from birth.
- transitional probabilities (how often one specified syllable precedes a second specified syllable). Computing these probabilities is more complex, but helps with learning words, because these probabilities reflect dependencies found in human language. Human infants as young as 7 months old can compute these probabilities from speech. But the researchers say that their use by animals has been demonstrated convincingly only in songbirds.
The researchers found that dogs can (like humans) learn to spot the beginning and end of words using both co-occurrence frequencies and transitional probabilities.
Which part of the dog brain?
The researchers also identified, using functional magnetic resonance (fMRI) measurements, two dog brain regions that play a role in spotting the start and end of words:
- The basal ganglia region. In humans, the same region is involved in processing sequences in general learning, and not just in learning language.
- The auditory cortex. In humans, this region supports specific learning about words.
The researchers say their findings show for the first time that a non-human animal (dogs) can learn to spot the beginning and end of words using complex computations (involving both co-occurrence frequencies and transitional probabilities). They do this in a way surprisingly similar to humans—and in similar regions of the brain.
The researchers reported that their results cannot show how these skills arose in dogs. Are they common to many mammals? Or did dogs develop these skills only by living in a language-rich environment over thousands of years of domestication by humans?
There is a very good and clear short video (2 minutes) summarising the research at Dogs learn about word boundaries as human infants do/ Curr. Biol., Oct. 29, 2021 (Vol. 31, Issue 24) – YouTube.
A non-technical summary is at Dogs learn about word boundaries as human inf | EurekAlert!
The paper itself is at Neural processes underlying statistical learning for speech segmentation in dogs: Current Biology (cell.com)