Maswao: An automatically collected, prepared, and benchmarked dataset for the South African Sign language

| Source: arXiv AI

Tags: South African Sign Language, Maswao, sign language translation, Transformer, automatic speech recognition, low-resource languages

Researchers built Maswao, a South African Sign Language dataset assembled automatically from news broadcasts, showing a low-labour route to data for under-served sign languages, though translation scores remain low.

Details

Only a few of the more than 200 recognized sign languages, such as American, British, and German, benefit from recent advances in sign language processing. This paper proposes a framework to collect and prepare datasets with minimal human labour, using South African Sign Language (SASL) as the test case. The pipeline pulls SASL video from public sources like news broadcasts, transcribes the accompanying audio with existing ASR tools, crops the signing area, and uses gradient descent to find low-motion valleys that mark sentence boundaries, producing a weakly sentence-aligned parallel corpus. On sign-to-text translation, Transformer models beat recurrent baselines, but the best reached only BLEU-4 of 1.54, ChrF 18.09, and ROUGE-L 10.51, so the task remains hard.