Right now, modern artificial intelligence and speech tools only seem to speak a handful of dominant global languages. Hundreds of local and regional tongues are left out entirely. A new project featured at the LREC 2026 conference is working to change that by expanding Mozilla's Common Voice framework across Pakistan.
Pakistan is home to more than 70 distinct languages, and nearly 30 of them are facing the threat of extinction. In the past, any digital or software development focused almost exclusively on Urdu and a couple of other major tongues. Because of this, millions of people who speak regional languages such as Saraiki cannot use voice search, speech-to-text programs, or automated translation apps.
To fix this neglect, a year-long community project funded by Mozilla set out to build open speech databases for 39 local Pakistani languages. Instead of using messy web scraping tools that usually bring in poor data and formatting errors, the team focused on gathering real, authentic material from the ground up. They used locally written texts, everyday conversational sentences that capture natural speech, and traditional folk songs and poetry to make sure the recordings truly reflect the culture and rhythm of the language.
For anyone studying linguistics or sentence structure, open speech databases like these are essential. Researchers need clean, natural audio recordings to properly map out grammar, word patterns, and speech sounds in under-documented languages. By sharing these datasets openly with the public, this project connects traditional field linguistics with modern machine learning, helping endangered languages survive and grow in the digital age.
