New AfriBED Dataset Addresses Speech Recognition Bias in Low-Resource African Languages

Researchers have developed AfriBED, a new dataset designed to evaluate and mitigate speaker-attribute bias in Automatic Speech Recognition (ASR) systems, specifically for low-resource African languages. While ASR technology has advanced significantly, existing systems often show performance disparities based on speaker demographics like gender and accent. This bias is particularly problematic for languages with limited digital resources, where such systems are already less accurate.

AfriBED focuses on two specific Zambian languages: Bemba and Nyanja. The dataset includes 9 hours of Bemba speech from 80 speakers and 18 hours of Nyanja speech from 127 speakers. This rich collection of comparable speech across different demographic subgroups is crucial for accurately assessing and understanding bias.

The initial experiments conducted using AfriBED have demonstrated its effectiveness in identifying gender and speaker-type bias within ASR systems targeting these languages. This development is a significant step towards creating more equitable and inclusive AI technologies for African linguistic communities, ensuring that ASR systems perform reliably for all users regardless of their demographic attributes.