English Speech Recognition Benchmark Expands to Egyptian and Nigerian Accents
Hugging Face has widened the accent coverage of its English speech recognition benchmark in partnership with Liva AI. Using private, unreleased data is meant to discourage competition driven purely by benchmark scores.
Hugging Face announced on November 6 that it is adding Egyptian and Nigerian English accent data to the Open ASR Leaderboard, its comparison table for English speech recognition performance. The material is private data obtained through a partnership with Liva AI. The addition gives a broader view of how a speech recognition model’s scores shift depending on which accent a speaker uses, even when everyone is speaking English.
This is not the first accent expansion. Hugging Face previously added 11 sets of private evaluation data supplied by Appen and DataOceanAI. Those sets cover both scripted read speech and free-form conversational speech across five accents. A dataset of Indian English was then added in collaboration with Voice Arena, and now two more regional accents join the mix.
Hugging Face said the goal is to reflect a wider range of accents in evaluation and to raise confidence in the results. By securing new private datasets and new data providers, the company aims to curb attempts to game benchmark scores. Accent classification follows Kachru’s three concentric circles theory, which divides English-speaking regions into three layers. If the number of accents grows further, the plan is to split some of the evaluation categories.
Image: huggingface.co
For anyone choosing an English speech recognition tool, the range of comparable options widens. Model recognition scores can now be checked for Egyptian and Nigerian English as well. What to watch next is how much further accent-by-accent coverage expands, and how precisely that lets users distinguish between models.