New AfriBED Dataset Addresses Speech Recognition Bias in Bemba and Nyanja Languages

Researchers have introduced AfriBED, a pioneering speech recognition bias evaluation dataset specifically designed for two low-resource African languages: Bemba and Nyanja. This initiative aims to address a critical shortcoming in current Automatic Speech Recognition (ASR) systems, which often exhibit performance biases against speakers based on demographic attributes like gender and accent.

The development of AfriBED is crucial because accurately identifying and quantifying these biases requires specialized datasets that offer comparable speech samples across different demographic subgroups. Without such resources, it's challenging to build ASR systems that perform equitably for all users, particularly in linguistic contexts where data is scarce.

The dataset for Bemba includes approximately 9 hours of speech from 80 speakers across 3928 utterances, while the Nyanja dataset comprises about 18 hours of speech from 127 speakers, totaling 6260 utterances. Initial experiments using AfriBED have already demonstrated its effectiveness in detecting and measuring gender and speaker-type biases within ASR systems for these specific languages.

This work represents a significant step towards more inclusive and fair AI technologies in Africa, providing a foundational tool for developers and researchers to build ASR systems that are less biased and more effective for African language speakers. By focusing on low-resource languages, AfriBED helps bridge the data gap that often marginalizes African linguistic diversity in AI development.