Your rare language corpus is missing for AI

Artificial intelligences are, by default, English-speaking: they have been fed from a web dominated by a handful of major languages. For dialects, regional languages, and sign languages, training data remains massively insufficient. And unlike English, there is almost nothing to retrieve online for a poorly written or digitized language: data must be produced, recorded, and then transcribed manually.

This effort has a cost, which is precisely what creates value. Transcribing one hour of audio in an underserved language can require around 30 to 36 hours of human work, whereas English requires only a fraction of that time (arXiv, 2025). In terms of price, quality annotated audio is around $90 to $180 per audio hour in English, with a 3 to 6 times premium for specialized or rare languages.

Concerned parties often unknowingly produce rare speech: multilingual or dialectal call centers, regional media and local radio stations, deaf associations and sign language interpreting services, but also the education sector, translation, and linguistic communities. What has value is the audio or video recording accompanied by its transcription: speech in a rare language or dialect, annotated sign language video, spontaneous speech from children, elders, or in the field.

Linguistic data involves people and communities: an ethical framework – explicit consent, anonymization, community respect – is not an option, and it is also what makes the data transferable and therefore valuable. The first concrete step is to determine if your corpus is valuable: launch a free assessment on d-nvest.