Sophia parle français

A Cajun French AI voice — built to help keep an endangered language alive. Trained on a single consumer laptop GPU, open and reproducible.

Severely
UNESCO: Louisiana French endangered
8 GB
trained on one laptop GPU
~8 hrs
Cajun French speech corpus
Open
method + recipe published

🎧 Hear Sophia speak Cajun French

Real output from the finetuned model — generated, not recorded. (Verified intelligible by automatic transcription.)

« Comment ça va, mon ami ? Ça fait longtemps que je t'ai pas vu, ouais. »
"How's it going, my friend? It's been a long time since I've seen you, yeah."
« Laissez les bons temps rouler ! On va faire un bon gombo ce soir, cher. »
"Let the good times roll! We're going to make a good gumbo tonight, dear."
« Mais regarde donc ça. La récolte de cannes est belle cette année, oui. »
"Well would you look at that. The cane harvest is beautiful this year, yes."

🗣️ Make Sophia say it (live)

Type a phrase in Louisiana/Cajun French and hear her speak it.

Why this matters

Louisiana French is classified by UNESCO as Severely Endangered, and Louisiana Creole (Kouri-Vini) as Critically Endangered (UNESCO Atlas of the World's Languages in Danger). Most fluent speakers are elderly; the language is largely not being passed to children.

A synthetic voice is not a replacement for living speakers — but it's a tool: for teaching, for narration, for accessibility, and to show that the language can live in modern technology. The point of building it on a single consumer laptop GPU is that the method stays in reach of community groups and classrooms, not just well-funded labs.

How it's built

A CosyVoice2 voice model finetuned on roughly eight hours of publicly available Cajun French speech (oral-history interviews, podcasts, and archival recordings), with the entire pipeline — data prep, training on 8 GB of VRAM, and synthesis — documented and open-source.

Open source github.com/Scottcjn/vintage-voice
Method reproducible recipe in docs/CAJUN_8H_FINETUNE.md

Built by an Opelousas native

This project is by Scott Boudreaux — an Opelousas native, of Acadian descent, building from Moss Bluff, Louisiana. Elyan Labs is exploring partnerships with Louisiana cultural and academic institutions to rebuild the voice on properly sourced, consented recordings.

Collaboration welcome archives, educators, and French/Creole community groups: scott@elyanlabs.ai