For much of the twentieth century, speaking French in a Louisiana schoolyard could get a child punished. The state's 1921 constitution mandated English-only instruction, and a generation grew up made to feel that their mother tongue was something shameful. The language that suffered was Cajun French — and now its would-be rescuers are turning to an unlikely tool: artificial intelligence.
A language pushed to the edge
Cajun French was carried to Louisiana by Acadians expelled from eastern Canada by the British in the 1750s and 1760s, and over two centuries it absorbed Spanish, Indigenous and Creole influences along the bayous of Acadiana. The English-only era left deep damage. By the time Louisiana created CODOFIL, the Council for the Development of French in Louisiana, in 1968 to try to reverse the decline, the language had retreated from public life.
Today an estimated 120,000 to 150,000 Louisianans still speak some form of Louisiana French, most of them elderly. Community educators warn that without intervention it could effectively disappear within a generation.
The machine that couldn't hear Cajun
The catalyst for one new effort was a small domestic absurdity. Joshua Caffery, director of the Center for Louisiana Studies at the University of Louisiana at Lafayette, asked a voice assistant to play a track by the late Cajun fiddler Dewey Balfa — and got a global pop star instead.
The mismatch points to a structural problem. Voice-recognition systems and language models are trained overwhelmingly on dominant languages and their standard forms. Louisiana French, with its distinctive sounds, older French roots and borrowed vocabulary, barely registers. An off-the-shelf system, the project notes, can hear the Louisiana French phrase meaning roughly "so, one fine day" and transcribe it as something else entirely — the meaning lost.
Building a dataset, one recording at a time
The university's response is a project called LaFLEUR — Louisiana AI for French Language Exploration, Understanding and Research. Its core work is painstaking: cutting decades of field recordings into short segments and pairing each with an expert-checked transcription, building the clean, reliable dataset that a model needs to learn from. Many recordings are old and noisy, layered with the background of kitchen-table music sessions, which trips up systems built for studio audio.
The aim is a transcription tool ordinary people could use on their own recordings — a grandchild capturing a grandparent's stories — without needing a linguistics degree to make sense of the result. The work sits alongside a growing field of research into how AI handles minority language varieties; a 2025 study at the City University of New York found that current speech-recognition systems can cause harm through bias when used with underrepresented varieties like Louisiana French.
Promise and limits
The broader movement to use AI for endangered languages is gaining ground worldwide, from Pacific languages to Indigenous tongues across the Americas, where the technology can speed up documentation that once took decades. But its advocates are clear about the limits. AI can transcribe a language; it cannot supply the young speakers, the immersion classrooms or the everyday use that actually keep one alive. There are subtler risks too: a dataset skewed toward one regional variant can quietly enshrine it over others, and digitizing a community's oral heritage raises real questions about consent and who controls the archive.
On the ground, organizers are pursuing the human work AI cannot do — including free Cajun French classes that have launched in communities across Acadiana, taught by people who have spent their lives inside the language. Cajun French survived Spanish rule, American annexation and a century of suppression. Whether it survives the digital age is a different test — and, in a place where its silence was once required by law, simply getting the machines to listen is a start.
