Early experimental results show that our approach improves model performance on all demographic groups in our evaluation datasets, with the largest gains in respect to accent inclusivity — demonstrating that improved fairness doesn’t mean a performance tradeoff. 2/3

→ View original post on X — @aiatmeta,