JUANA SUMMERS, HOST:
When you're listening to a familiar language, you hear individual words, but what if the language is unfamiliar?
UNIDENTIFIED PERSON #1: (Speaking Japanese).
SUMMERS: Unless you speak Japanese, that probably sounded like an unbroken stream of vowels and consonants. NPR's Jon Hamilton reports on what scientists are learning about how the brain extracts words from speech.
JON HAMILTON, BYLINE: Words are easy to recognize when they're written. A visible space shows where one word ends and another begins. Dr. Eddie Chang of the University of California, San Francisco says spoken language doesn't have that luxury.
EDDIE CHANG: One of the reasons why a foreign language sounds so fast is that you can't hear the pauses between words.
HAMILTON: Because most of the time there aren't any. Each syllable just bumps into the next. Yet Chang says in a familiar language, the brain creates the perception that we are hearing separate words.
CHANG: It's a really incredible automatic process that all of us have as human listeners.
HAMILTON: To learn more about that process, Chang and a team studied several dozen people about to have surgery for epilepsy. The participants already had electrodes in their brains, so the team was able to monitor neurons that respond to speech sounds. Chang says they found a distinct pattern.
CHANG: Your brain is processing this continuous stream of speech, but at every word boundary, the actual neural activity drops. There's a trough in the activity.
HAMILTON: As if the brain is putting a space between each word. To test this idea, Chang's team had participants listen to a sound loop.
UNIDENTIFIED PERSON #2: Boater, boater, boater, boater, boater (ph)....
CHANG: It's one stimulus that's just repeated over and over and over again. But your brain kind of forces you to hear one word or another.
UNIDENTIFIED PERSON #2: Turbo, turbo, turbo, turbo, turbo, turbo, turbo, turbo (ph)...
HAMILTON: You hear either boater or turbo depending on where your brain is placing the word boundary. And Chang says, in most people, that boundary will shift.
CHANG: Listeners typically will hear one word for a couple of seconds, and then their perception will switch to the other word.
HAMILTON: At the same exact moment their brain makes a new word boundary. Chang says the brain constantly tries to predict what it will hear next. An English-speaking brain, for example, knows that the sound bo could become boater or boba, but not baseball or Parcheesi. It also knows that the words peanut and butter are likely to lead to jelly. But Chang says prediction relies on a deep understanding of a language.
CHANG: We found that this encoding of words actually is a very specific process to one's mother tongue, to one's native language.
HAMILTON: Or languages, if you're multilingual. Nima Mesgarani of Columbia University says Chang's research helps explain how human brains are able to decode speech so quickly and accurately.
NIMA MESGARANI: They are showing that brain almost creates its own punctuation.
HAMILTON: Which helps it quickly assemble speech sounds into words and then sentences. Mesgarani says the latest artificial intelligence systems appear to use a similar strategy. He says, instead of following rules, these machines learn about speech the same way a child does - by listening.
MESGARANI: We basically just say, this is the input. This is what the desired output. And you have millions of hours of data, so learn it.
HAMILTON: Mesgarani says artificial brains, like their human counterparts, eventually learn to recognize where one word ends and another begins. He says this reflects something scientists are seeing in other areas of artificial intelligence.
MESGARANI: If you look at many different sort of AI models, the better they are at what they do, the more they align with the brain.
HAMILTON: These AI systems even share a key limitation with human brains - they can only identify word boundaries in a language they know really well. Jon Hamilton, NPR News. Transcript provided by NPR, Copyright NPR.
NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.
