Evaluating Automatic Speech Recognition Performance Across African English Accents
Domain:
natural language processingRecord type:
Automatic speech recognition (ASR) systems are widely used in digital assistants, transcription tools, and accessibility technologies. However, their performance often varies significantly depending on the speaker’s accent, particularly for underrepresented varieties of English. This study evaluates the performance of multiple ASR systems across a diverse set of African English accents.
Using speech samples collected from speakers across several African regions, the study measures transcription accuracy using word error rate (WER) and examines recurring patterns of misrecognition. Results show that ASR systems exhibit significantly higher error rates for underrepresented accents, highlighting systematic disparities in speech recognition technology.
The findings contribute to the growing body of research on algorithmic bias in speech technology and emphasize the need for more inclusive datasets and accent-aware model development. Recommendations are provided for improving ASR performance for African English speakers and for guiding future research in equitable speech recognition systems.