The Machine That Learns to Listen

Ask one of the world's leading chatbots to hold a conversation in isiZulu, and you will witness something between a performance and a guess. It will produce fluent-looking sentences that a first-language speaker recognises as subtly, persistently wrong: a wrong noun class here, a tense that does not exist there, an idiom translated so literally it collapses. The model is not lying. It simply was not built to listen to this voice. It learned the world from an internet that is overwhelmingly English, Mandarin and a handful of European tongues, and Africa's roughly two thousand languages barely registered in the data.

This is the polite name for a hard problem: low-resource languages. In machine-learning terms, a language is a resource, measured in how much clean, digitised text and speech exists to train on. By that cruel accounting, Yoruba, with tens of millions of speakers, is poorer than Welsh. The gap is not about how many people speak a language. It is about how much of that speech ever made it onto a server in a form a machine could swallow.