Answers are grounded in this paper and cite their sources.
Besides English, ten (10) indigenous languages are official languages in South Africa (SA). While this promotes and preserves languages, it easily creates communication barriers in daily activities, particularly when individuals relocate to different provinces for work or in religious settings. Immigrants and tourists can also easily encounter difficulty in communicating with indigenes. Regarding computational technologies, SA languages are generally under-resourced. This paper discusses the design of a bi-directional translation system that enables real-time communication between individuals speaking different SA languages. Existing translation systems are typically unidirectional, lack support for SA languages, and/or are not designed for long sentences or real-time communication. Our solution performs two-way translation between languages like a human interpreter does between two parties who do not understand each other’s language. Our proposed system implements a peer-to- peer network approach comprising speech recognition, deep-learning-based translation module, and voice synthesis. We implemented this cost-effective system on a regular laptop. Word error rates (WER) for English, isiZulu, and isiXhosa were 5.2%, 8.7%, and 6.1%, respectively. Text-to-text translation accuracy (BLEU score) between languages ranges between 0.75 and 0.82, with a perfect match (score of 1) returned when the source and destination languages are selected to be the same. Voice-to-voice translations were evaluated subjectively by employing the judgment of human speakers of the languages. Average performances (scale of 1-10) were 8 (isiZulu to isiXhosa), 6 (isiZulu to English), 7 (isiXhosa to isiZulu), 8 (English to isiZulu), 7 (English to isiXhosa). Translation latency was consistently below 350ms, which shows that real-time translation is satisfactorily achieved.
No takes yet. Share an insight, caveat, or question.
Simelane et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: