IACC-HuBERT cuts dysarthric speech WER to 21.0% from 28.9%

IACC-HuBERT builds on HuBERT, a self-supervised model that learns representations from speech. The paper's authors, Paban Sapkota and Hemant Kumar Kathania, argue that representations learned mostly from healthy speech may not suit every dysarthric speaker, so they condition the features on an intelligibility embedding extracted from the speech itself.

The method trains an intelligibility classifier on utterance-level HuBERT features, then uses the 128-dimensional representation from its penultimate layer as the conditioning vector. FiLM and Gated-FiLM modules then scale and shift HuBERT's hidden dimensions. The frontend stays frozen except for the added modules, which have 6.3M trainable parameters for FiLM-only and 15.78M for Gated-FiLM.

On TORGO, the approach improved over plain HuBERT for most speakers, but not all. The authors note that the gate's benefit varies by speaker and that aggregation details need checking, so the result is a promising direction for adapting pretrained representations rather than a uniform fix.