- Professor Joshua Caffery developed an AI voice recognition model for Louisiana French after Alexa played Dua Lipa’s music instead of Cajun artist Dewey Balfa, showing that mainstream AI does not yet understand local languages.
- The project utilizes nursery rhymes, folk recordings, interviews, and historical performances to train an automatic speech recognition model specifically for Louisiana French.
- The Center for Louisiana Studies archives more than 12,000 hours of audio materials, including oral histories, folk music, and recordings across multiple formats spanning many decades.
- Researchers built a gold-standard dataset through manual transcription, then fine-tuned an open-source model named Tataille to improve accuracy.
- The post-trained model recognizes the distinctive spelling and verb conjugations characteristic of Louisiana French, rather than inferring from mainstream data on the Internet.
- Experts warn that AI misidentifying minority languages could impact recruitment, healthcare, and various other digital services.
- Louisiana French was once a widespread language, but after the 1921 Constitution, many families stopped passing it down due to fears of discrimination. In 2023, approximately 120,000 speakers remained.
- The research team hopes that AI will help preserve, teach, and encourage new generations to create music, culture, and digital content in their local language.
- 📌 Conclusion: Generative AI and speech recognition serve not only productivity but can also become vital tools for cultural heritage preservation. By training the model on over 12,000 hours of archival materials and local data, the research team is bringing Louisiana French into the digital era. This approach also opens up opportunities to protect many other minority languages, helping communities retain control over their identity and digital assets in the future.
