Lelapa AI, a new artificial intelligence lab based in South Africa, is set to launch its first natural language model solution, Vulavula, in the coming months. The start-up aims to build AI-based products and services that solve African problems.
Vulavula will apply entity recognition, speech-to-text and text-to-speech technologies to under-represented languages, specifically South African ones. Language models like ChatGPT may be wowing us all, but they generally have limited capabilities for low-resourced languages. Research cited by Lelapa has found that many multilingual language models are trained on 'dangerous, offensive and frankly garbage data.' The models don't perform well on many languages, especially African ones.
Meanwhile, 90% of the internet is in English, but only one in 10 South Africans speak English at home. Vulavula can address that by offering automatically generated subtitles for videos or translation for B2C chat interfaces. While primarily available in isiZulu for now, Lelapa's team has done much groundwork to support more languages, including Xhosa, Sesotho, Setswana and Afrikaans, potentially allowing up to 80% of the population to be able to speak and be spoken to in their own languages.
