Scientists have made new improvements to a “brain decoder” that uses artificial intelligence (AI) to convert thoughts into text.

Their new converter algorithm can quickly train an existing decoder on another person’s brain, the team reported in a new study. The findings could one day support people with aphasia, a brain disorder that affects a person’s ability to communicate, the scientists said.

A brain decoder uses machine learning to translate a person’s thoughts into text, based on their brain’s responses to stories they’ve listened to. However, past iterations of the decoder required participants to listen to stories inside an MRI machine for many hours, and these decoders worked only for the individuals they were trained on.

“People with aphasia oftentimes have some trouble understanding language as well as producing language,” said study co-author Alexander Huth, a computational neuroscientist at the University of Texas at Austin (UT Austin). “So if that’s the case, then we might not be able to build models for their brain at all by watching how their brain responds to stories they listen to.”

In the new research, published Feb. 6 in the journal Current Biology, Huth and co-author Jerry Tang, a graduate student at UT Austin investigated how they might overcome this limitation. “In this study, we were asking, can we do things differently?” he said. “Can we essentially transfer a decoder that we built for one person’s brain to another person’s brain?”

The researchers first trained the brain decoder on a few reference participants the long way — by collecting functional MRI data while the participants listened to 10 hours of radio stories.

Then, they trained two converter algorithms on the reference participants and on a different set of “goal” participants: one using data collected while the participants spent 70 minutes listening to radio stories, and the other while they spent 70 minutes watching silent Pixar short films unrelated to the radio stories.

Using a technique called functional alignment, the team mapped out how the reference and goal participants’ brains responded to the same audio or film stories. They used that information to train the decoder to work with the goal participants’ brains, without needing to collect multiple hours of training data.

Next, the team tested the decoders using a short story that none of the participants had heard before. Although the decoder’s predictions were slightly more accurate for the original reference participants than for the ones who used the converters, the words it predicted from each participant’s brain scans were still semantically related to those used in the test story.

For example, a section of the test story included someone discussing a job they didn’t enjoy, saying “I’m a waitress at an ice cream parlor. So, um, that’s not … I don’t know where I want to be but I know it’s not that.” The decoder using the converter algorithm trained on film data predicted: “I was at a job I thought was boring. I had to take orders and I did not like them so I worked on them every day.” Not an exact match — the decoder doesn’t read out the exact sounds people heard, Huth said — but the ideas are related.

“The really surprising and cool thing was that we can do this even not using language data,” Huth told Live Science. “So we can have data that we collect just while somebody’s watching silent videos, and then we can use that to build this language decoder for their brain.”

Using the video-based converters to transfer existing decoders to people with aphasia may help them express their thoughts, the researchers said. It also reveals some overlap between the ways humans represent ideas from language and from visual narratives in the brain.

“This study suggests that there’s some semantic representation which does not care from which modality it comes,” Yukiyasu Kamitani, a computational neuroscientist at Kyoto University who was not involved in the study, told Live Science. In other words, it helps reveal how the brain represents certain concepts in the same way, even when they’re presented in different formats.

The team’s next steps are to test the converter on participants with aphasia and “build an interface that would help them generate language that they want to generate,” Huth said.

Scientists have made new improvements to a “brain decoder” that uses artificial intelligence (AI) to convert thoughts into text.

Their new converter algorithm can quickly train an existing decoder on another person’s brain, the team reported in a new study. The findings could one day support people with aphasia, a brain disorder that affects a person’s ability to communicate, the scientists said.

A brain decoder uses machine learning to translate a person’s thoughts into text, based on their brain’s responses to stories they’ve listened to. However, past iterations of the decoder required participants to listen to stories inside an MRI machine for many hours, and these decoders worked only for the individuals they were trained on.

“People with aphasia oftentimes have some trouble understanding language as well as producing language,” said study co-author Alexander Huth, a computational neuroscientist at the University of Texas at Austin (UT Austin). “So if that’s the case, then we might not be able to build models for their brain at all by watching how their brain responds to stories they listen to.”

In the new research, published Feb. 6 in the journal Current Biology, Huth and co-author Jerry Tang, a graduate student at UT Austin investigated how they might overcome this limitation. “In this study, we were asking, can we do things differently?” he said. “Can we essentially transfer a decoder that we built for one person’s brain to another person’s brain?”

The researchers first trained the brain decoder on a few reference participants the long way — by collecting functional MRI data while the participants listened to 10 hours of radio stories.

Then, they trained two converter algorithms on the reference participants and on a different set of “goal” participants: one using data collected while the participants spent 70 minutes listening to radio stories, and the other while they spent 70 minutes watching silent Pixar short films unrelated to the radio stories.

Using a technique called functional alignment, the team mapped out how the reference and goal participants’ brains responded to the same audio or film stories. They used that information to train the decoder to work with the goal participants’ brains, without needing to collect multiple hours of training data.

Next, the team tested the decoders using a short story that none of the participants had heard before. Although the decoder’s predictions were slightly more accurate for the original reference participants than for the ones who used the converters, the words it predicted from each participant’s brain scans were still semantically related to those used in the test story.

For example, a section of the test story included someone discussing a job they didn’t enjoy, saying “I’m a waitress at an ice cream parlor. So, um, that’s not … I don’t know where I want to be but I know it’s not that.” The decoder using the converter algorithm trained on film data predicted: “I was at a job I thought was boring. I had to take orders and I did not like them so I worked on them every day.” Not an exact match — the decoder doesn’t read out the exact sounds people heard, Huth said — but the ideas are related.

“The really surprising and cool thing was that we can do this even not using language data,” Huth told Live Science. “So we can have data that we collect just while somebody’s watching silent videos, and then we can use that to build this language decoder for their brain.”

Using the video-based converters to transfer existing decoders to people with aphasia may help them express their thoughts, the researchers said. It also reveals some overlap between the ways humans represent ideas from language and from visual narratives in the brain.

“This study suggests that there’s some semantic representation which does not care from which modality it comes,” Yukiyasu Kamitani, a computational neuroscientist at Kyoto University who was not involved in the study, told Live Science. In other words, it helps reveal how the brain represents certain concepts in the same way, even when they’re presented in different formats.

The team’s next steps are to test the converter on participants with aphasia and “build an interface that would help them generate language that they want to generate,” Huth said.

Artificial intelligence has transformed many industries, and transcription is no exception. While AI transcription software offers un-human speed, it’s far from perfect. Anyone who has used these tools knows they can produce some truly baffling results. These moments of machine misinterpretation are not just amusing but also a reminder of why human transcription remains essential.

Recently, while doing some transcription work, we’ve encountered some examples of AI transcription gone hilariously wrong. From bizarre substitutions to completely nonsensical sentences, these errors highlight the limitations of relying solely on algorithms to understand HUMAN language. Let’s dive into a few of these blunders that show why the human touch is still irreplaceable in transcription.

So we took off in convoy back to the Suez Canal, through the Suez Canal, back to Missouri Bizerte.

I guess you’ve heard of Anahita Enewetak.

My father, his name was Mokosak Markus Zack.

Did you find out when he was transported to Terezin, stat Theresienstadt?

We had a tick tock to tiptoe all the way back to the Philippines.

And the chaplain at Meredith’s that married us was a Catholic chaplain who was from South Portland, Maine.

So, anyway, then when they came with the draft, as I said, I was a for declassification 4D classification.

And this dwarf And Düsseldorf was just like some of the pictures I saw here.

We were at a village called Wingen sur Moder. Wingen on the motor River. Wingen-sur-Moder—Wingen on the Moder River.

They were bored with bartered everything for a piece of food.

No, Mr. Battle fatigue, Pietroforte, don’t do that.

And we’re just biding our time before the attack on Hawken Aachen.

It’s similar to a picture that I have for Michelle Hall from a shell hole.

Artificial intelligence has transformed many industries, and transcription is no exception. While AI transcription software offers un-human speed, it’s far from perfect. Anyone who has used these tools knows they can produce some truly baffling results. These moments of machine misinterpretation are not just amusing but also a reminder of why human transcription remains essential.

Recently, while doing some transcription work, we’ve encountered some examples of AI transcription gone hilariously wrong. From bizarre substitutions to completely nonsensical sentences, these errors highlight the limitations of relying solely on algorithms to understand HUMAN language. Let’s dive into a few of these blunders that show why the human touch is still irreplaceable in transcription.

So we took off in convoy back to the Suez Canal, through the Suez Canal, back to Missouri Bizerte.

I guess you’ve heard of Anahita Enewetak.

My father, his name was Mokosak Markus Zack.

Did you find out when he was transported to Terezin, stat Theresienstadt?

We had a tick tock to tiptoe all the way back to the Philippines.

And the chaplain at Meredith’s that married us was a Catholic chaplain who was from South Portland, Maine.

So, anyway, then when they came with the draft, as I said, I was a for declassification 4D classification.

And this dwarf And Düsseldorf was just like some of the pictures I saw here.

We were at a village called Wingen sur Moder. Wingen on the motor River. Wingen-sur-Moder—Wingen on the Moder River.

They were bored with bartered everything for a piece of food.

No, Mr. Battle fatigue, Pietroforte, don’t do that.

And we’re just biding our time before the attack on Hawken Aachen.

It’s similar to a picture that I have for Michelle Hall from a shell hole.

Q: I was surprised when autocorrect changed “intermittent” to “intermit.” I checked and, lo and behold, there is a word “intermit.” Does it not strike you as odd that the base-form is less known than its “built-up” version?

A: We don’t use, or recommend using, the autocorrect function in a word processor. Our spell-checkers flag possible misspellings but don’t automatically “fix” them. Word processors have dictionaries, but not common sense—at least not yet!

As for the words you’re asking about, the adjective “intermittent” (irregular or occurring at intervals) is indeed more common than the verb “intermit” (to suspend or stop). In fact, the verb barely registered when we compared the terms on Google’s Ngram Viewer.

However, “intermittent” isn’t derived from “intermit,” though both ultimately come from different forms of the Latin verb intermittere (to interrupt, leave a gap, suspend, or stop), according to the Oxford English Dictionary. The Latin verb combines inter (between) and mittere (to send, let go, put).

When “intermit” first appeared in English in the mid-16th century, it meant to interrupt someone or something, a sense the OED describes as obsolete.

The modern sense of the verb—“to leave off, give over, discontinue (an action, practice, etc.) for a time; to suspend”—showed up in the late 16th century.

It means “leave off” in the dictionary’s earliest citation for the modern usage: “Occasions of intermitting the writing of letters” (from A Panoplie of Epistles, 1576, by Abraham Fleming, an author, editor, and Anglican clergyman).

As we’ve said, “intermit” isn’t seen much nowadays. English speakers are more likely to use other verbs with similar senses, such as “cease,” “quit,” “stop,” “discontinue,” “interrupt,” or “suspend.”

When the adjective “intermittent” appeared in the early 17th century, Oxford says, it described a medical condition such as a pulse, fever, or cramp “coming at intervals; operating by fits and starts.”

The earliest OED citation is from an English translation of Plutarch’s Ἠθικά (Ethica, Ethics), commonly known by its Latin title Moralia (The Morals), a collection of essays and speeches originally published in Greek around the end of the first century:

“Beating within the arteries here and there disorderly, and now and then like intermittent pulses” (from The Philosophie, Commonly Called, The Morals, 1603, translated by Philemon Holland).

The adjective later took on several other technical senses involving irregular movement, but we’ll skip to its use in everyday English to mean occurring at irregular intervals. The earliest OED citation for this “general use” is expanded here:

“Northfleet a disunited Village of 3 Furlongs, with an intermittent Market on Tuesdays, from Easter till Whitsuntide only” (Britannia, or, An illustration of the Kingdom of England and Dominion of Wales, 1675, by the Scottish geographer John Ogilby).