SALSA: Speech Aware LLM Adaptation via Learned Steering Activation Vectors

In three linesSALSA adapts speech-aware LLMs via layer-wise learned steering vectors optimized with supervised objectives. Tested on children's speech, multilingual, and Mandarin-English code-switching, it achieves up to 46.8% relative improvement over zero-shot. Steering the encoder's later layers outperforms steering the LLM backbone.

Summary generated by Claude — human-verified