Learning task-specific subspaces via interventional post-training of speech foundation models

In three linesPost-training refinement method for speech foundation models using interventional contrastive learning. Transforms entangled representations into separate content and speaker subspaces via interventional dataset and multi-part contrastive loss. Improves out-of-domain speaker verification and keyword spotting performance.

Summary generated by Claude — human-verified