Routed Sparse Autoencoders Disentangle Speech Linguistics and Paralinguistics

Using TopK sparse autoencoders with cross-factor adversaries, this method separates linguistic and paralinguistic information in SPEAR and WavLM encoders. The approach enables feature-space interventions to swap factors like emotion or prosody while preserving linguistic content without retraining.