Current progress and boundaries
Machine translation and speech recognition have advanced for major languages, but many local languages still lack high-quality, properly licensed training and evaluation materials.
Why it remains difficult
Dialect differences, spoken language, and cultural context can make literal translations misleading. Data from small communities can be over-collected and may not be sufficient to maintain models reliably.
Paths worth exploring
Let teachers, language users, and communities decide how data may be used. Use revocable consent, openly available evaluation sets, and local maintenance, prioritizing real classroom needs.
How to evaluate success
Have native speakers assess meaning, tone, and cultural appropriateness. Also measure teacher preparation burden, student comprehension, and community satisfaction with data governance.