South Africa's multilingual reality can get lost when local languages present a challenge for developing language technology. Venelize de Lange from media update considers what it means when the language we use does not fit the systems built to understand it.
Ever heard of vernaculars or code-switching? We use them all the time as South Africans, throwing in a "lekker" whenever we get the chance, or a "Mzansi" when speaking about our beloved country — sometimes even borrowing another language's structures just to land a sentence right.
South Africa is rich in culture, meaning it seeps in everywhere and in everything, including language. While that makes for a vibrant country full of different flavours and tastes, it creates a problem for uniform systems, especially as it pertains to language technology.
Ask a transcription tool to capture a South African conversation, and its limitations show up quickly: Afrikaans gets mistaken for Dutch, isiZulu becomes muddled when speakers switch between languages mid-sentence, and translation tools don't always account for cultural context.
This becomes counterproductive when organisations are trying to understand what their audiences are saying — how and what they are communicating — as conversation cannot be analysed if its cultural nuance is stripped away.
Part of the problem is structural. Training data is concentrated around languages with large quantities of text available online, predominantly in English and other European and Chinese languages. Many African languages, by contrast, are primarily spoken rather than written, leaving less text from which language systems can learn.
In South Africa specifically, English dominance in AI and tech development locks out the majority of the population who speak an indigenous language as their first tongue, especially if they have limited understanding of English, or none at all.
So, English becomes the default — but only because it's the language these systems understand best. However, language technology that consistently rewards standardised English and battles with vernacular language, code-switching and local expressions runs the risk of quietly pressuring people to adjust their language to fit the system.
University of Pretoria Professor Vukosi Marivate, who worked on a major African-language AI dataset, puts it simply, "We think in our own languages, dream in them and interpret the world through them." If the technology doesn't reflect that, he warns, a whole group of people risk being left behind.
Language carries culture — but language is also a relationship. Projects like African Next Voices, which has recorded thousands of hours of speech across 18 African languages, show that the influence can run both ways: real speech, gathered from real communities, can reshape the tools meant to understand it.
Without that kind of investment, though, the risk runs the other way — indigenous languages left out of AI development face marginalisation, and a quiet homogenisation takes root.
Did you find this article insightful? Let us know in the comments section below.
*Image courtesy of Canva
**Information sourced from iAfrica, BBC and Wits Vuvuzela