CAFE Method Mitigates English Performance Loss in Multilingual ASR
October 15, 2026
CAFE uses language-routed adapters to confine Centered Kernel Alignment (CKA) feature distillation to non-English parameters. This prevents CKA gradients from overlapping with English-dominated subspaces, maintaining high English accuracy while improving non-English word error rates.
HOW THIS AFFECTS YOU
●
researcherYou can use adapter-based confinement to solve the trade-off between multilingual capability and high-resource language performance.