A painting by Andrew Hopkins titled “Louisiana French Afro-Creole, Grandmother’s Wisdom” (2026).Credit...Andrew Hopkins
While relaxing a couple of years ago, Prof. Joshua Caffery found himself in the mood to unwind with some old-time Cajun music. He asked Amazon’s Alexa to play selections from Dewey Balfa, a celebrated fiddler and singer credited with popularizing the genre.
Instead, Alexa frustratingly steered him to the catalog of the modern pop artist, Dua Lipa, Caffery said.
“I love Dua Lipa,” said Caffery, the director of the Center for Louisiana Studies at the University of Louisiana at Lafayette. “Don’t get me wrong. But it seems problematic if you’re interested in a different kind of culture and you want to surround yourself with the music of your region. That, to some degree, is threatening my hold on these things I love.”
Louisiana French, the oral dialect of which Balfa was a cultural guardian, is part of the Bayou’s societal DNA, a link to its history, music and identity. Today, Caffery described the language as struggling and endangered, a notion reinforced by Alexa’s overlooking Balfa.
In response, Caffery assembled a small team at the center to train its own language learning model in automatic speech recognition for Louisiana French, drawing from a trove of historical artifacts and interviews.
Over the months, as the learning language model is trained on bits of the language — such as an old-age French nursery rhyme — it brings centuries-old dialect closer into the digital age.
“It’s scary how fast all this happened,” Caffery said. “I don’t even think I knew a large-language model existed two years ago. Now, they’re everywhere.”
The efforts are part of inroads to preserve dialects at risk of being lost, and bridge a disconnect between large-language model machines, while maintaining a community’s ability to control and own its digital destiny.
In the United States, questions exist over how to preserve languages like Louisiana French, the Gullah Geechee language(the English-based Creole language of some coastal regions in the Atlantic southeast) and Appalachian English in the modern, digital age.
Globally, researchers at the University of Edinburgh are using artificial intelligence to strengthen and revive Scottish Gaelic and Manx. Google recently announced the availability of voice recordings of nearly 30 sub-Saharan African languages to help build tools like voice assistants and translation apps.
The consequences of the language gap can be far greater than an A.I. assistant confusing musical artists, said Christine Mallinson, a professor of language, literacy, and culture at the University of Maryland, Baltimore County.
The importance of accurate speech recognition becomes greater as important tasks like job hiring and medical transcriptions become more automated and digitized, she said.
“Social differences are encoded in language,” Mallinson said. “There’s accents, patterns of grammar, word choice. Those differences are connected to our families, our neighborhoods, our age and gender and racial and ethnic and cultural backgrounds and where we grew up.
“If A.I. speech systems make more errors for speakers of underrepresented languages or language varieties, then there can be these serious downstream consequences,” she continued.
For centuries, Louisiana French was the predominant language spoken in South Louisiana. In 1921, a new state constitution declared English the primary language. Many parents stopped teaching their children the language out of fear of discrimination, as students who spoke Louisiana French in class were often punished with knuckle-rappings.
A reversal came in 1968 when the Council for the Development of French in Louisiana was created to advance French, largely through education and community initiatives. In 2023, the Advocate of Baton Rouge estimated about 120,000 Louisianans still spoke French.
Caffery grew up in Franklin, La., an antebellum town on Bayou Teche. As a child, he ate Cajun and Creole food. His grandparents sometimes sang Old French songs and passed along phrases in the language.
“Whether you spoke it or speak it, the language is floating in the air,” he said. “There’s this feeling of there is this beautiful thing that we want to hang on to.”
His work to do just that is painstakingly detailed.
The Center of Louisiana Studies houses a vast trove of Cajun and Creole folklore recordings that include over 12,000 hours of oral histories, field recordings and music performances recorded on everything from wax cylinders to reel-to-reel tape. Barry Jean Ancelet, a professor emeritus at the University of Louisiana at Lafayette, is a cultural activist who has donated much of his recording collection, which dates back to the 1970s.
An interviewee’s remembrance of the locally infamous 1896 murder of a Louisiana merchant. This recording was used to test the A.I. model’s ability to decode regional speech.
The language is almost entirely preserved in vocal recordings. “A lot of it is sung in the language, and it’s something that’s vanishing,” Caffery said. “Just in the same way you’d want food that’s local to your region, it’s important to have culture that is local to the region that makes you feel rooted and certain.”
Initially, Caffery and his team fed some of the Louisiana French recordings into an automated speech recognition model trained in standard English. The results, he said, were better than anticipated but contained enough errors to render them unusable.
To reduce the flaws, the linguists Amanda LaFleur and Colby LeJeune have transcribed audio from the songs and interviews to build what they term as a ground-truth data set for the artificial intelligence tool to learn from.
Peyton Leathem-Boe, a research scientist, is working with the open-source model they have named the “tataille,” a monster or boogeyman in Louisiana French.
On a recent weekday, the group met on the second floor of the center, housed inside the J. Arthur Roy House, a renovated, white-frame Queen Anne mansion.
On a large monitor, Leathem-Boe displayed side-by-side comparisons taken from a snippet of Sam and Jesse Stafford signing “Trois Jolis Tambours,” a classic French nursery rhyme, nearly a century ago. On the left, the screen showed the outcome with the help of LaFleur and LeJeune’s transcriptions. The right depicted the results without the model.
“Oh, what’s that?” LaFleur asked, trying to decipher a portion of the text on the untrained model. “It’s not Hebrew or Arabic.”
“Definitely not in Arabic.” LeJeune said.
“Oh, it’s Hindi,” LaFleur said.
“That’s one of the interesting things is that if you give it, say, really chunky, really crunchy audio, or if you give it music, it will just generate random characters and it will predict that that is that language,” Leathem-Boe said.
The variations in the two transcriptions caught LaFleur’s attention. LaFleur turned to how the trained model conjugated the past tense of “to have” in Louisiana French.
“It actually came up with this spelling which is uniquely Louisianian,” LaFleur said. “It recognized it as something that we gave it. It didn’t get it out of ChatGPT or all the other millions of data. That’s us. So that’s exciting.”
The trained model was not 100-percent accurate, but a marked improvement over the standard model. The group will input what the model missed, fine tune it and feed it back again.
“If you train it enough on the right information, it will theoretically eventually do it the right way,” Caffery said.
Caffery is also a musician who performed on Grammy-nominated albums based in Louisiana music. He hopes the model will eventually allow musicians who are not fluent in Louisiana French to read lyrics and become inspired.
And yes, he still hopes to one day ask Alexa for Dewey Balfa and not hear Dua Lipa.
“A.I. is potentially a very powerful tool in our arsenal to try to preserve and galvanize and stimulate tradition,” he said. “Not tradition as a static thing of just preserving old things, but making sure that the gears of tradition remain lubricated and that people keep creating within the tradition.”
From the comments
Trois Jolis Tambours recording: John A. and Alan Lomax, Library of Congress. Martin Begnaud recording: Barry Jean Ancelet, Center for Louisiana Studies.
Jonathan Abrams is a Times reporter who writes about the intersections of sports and culture and the changing cultural scenes in the South.