Someone asked Amazon's Alexa to play Dewey Balfa, the celebrated Cajun fiddler credited with popularizing Louisiana French music. Alexa served up Dua Lipa instead.

That's not just a funny tech fail. It's a measurement. When your language isn't in the training data, the algorithm acts like you don't exist.

The Language South Louisiana Spoke for Centuries

Louisiana French was the predominant language of South Louisiana for centuries. Not a folk dialect. Not a curiosity. The actual language millions of people used for daily life, love letters, and lullabies.

In 2023, the Advocate of Baton Rouge estimated about 120,000 Louisianans still spoke it. Which sounds substantial until you consider how few that is compared to where it started, and how much faster language loss tends to move than recovery.

Linguists describe it as struggling and endangered. That's the careful word for "we might lose this."

How You Kill a Language

Step one: write a constitution. In 1921, Louisiana declared English the primary language. One document, sweeping consequences.

Step two: make the language painful. Students who spoke Louisiana French in class got their knuckles rapped. Many parents stopped teaching their children the language out of fear of discrimination. Not because they wanted to. Because they were afraid of what it might cost their kids.

Step three: wait.

Louisiana created the Council for the Development of French in Louisiana in 1968, specifically to push back through education and community programs. A whole government body trying to undo what another government document started forty-seven years earlier. Language loss is fast. Recovery is slow.

Teaching a Machine to Hear Nursery Rhymes

Professor Joshua Caffery, director of the Center for Louisiana Studies at the University of Louisiana at Lafayette, is running a different kind of recovery project.

His team is training a speech recognition model in Louisiana French. The kind of AI that powers Alexa and Siri, except this one is trained on historical artifacts and interviews, including old Louisiana French nursery rhymes.

That last part matters more than it sounds. Nursery rhymes are where a language lives at its most fundamental. Rhythm, vowel sounds, the patterns children absorb before they can read. Training data pulled from nursery rhymes isn't nostalgic. It's structural. It's the phonological core of a language captured while it was still being passed between generations.

Researchers like Christine Mallinson, a professor of language, literacy, and culture at the University of Maryland, Baltimore County, work in the field where language and community identity intersect. The urgency is real: once the last fluent speakers are gone, transcripts and recordings are all that remain. A living language can't be reconstructed from paper alone.

What the Algorithm Gets Wrong

Dewey Balfa spent his career making sure people could hear Louisiana French music. He's one of the most celebrated figures in Cajun culture. Ask Alexa about him and you get a British pop star.

That's not an insult to Dua Lipa. It's a gap in the data. The algorithm knows what it's been fed. Louisiana French wasn't in the menu.

Caffery's team is working to change that. Training a speech model on the actual sounds of the language, pulled from artifacts that captured it while it was alive in communities, is a way of making the language legible to systems that currently can't find it.

One possibility is that the nursery rhymes become seed data for much larger models eventually. That the old lullabies end up as the foundation for a language getting a second life in the systems that shape what we hear.

120,000 Speakers Is a Number Worth Sitting With

Languages don't die all at once. They die one conversation at a time, when the default choice is the safe language, the school language, the one that won't get your knuckles rapped.

Every word game you love is a record of a language that survived. The strange letter combinations in Scrabble dictionaries, the etymology trails in crosswords, the borrowed words that show up in Wordle. Languages that are thriving leave fingerprints everywhere.

Languages that are struggling go quiet in the systems that decide what we find.

The algorithm confused Dewey Balfa with Dua Lipa. But it's learning. That's the whole point. That's something.

Source: Languagehat