Activation Steering Enables Emotion Control in Moshi Speech Models
October 8, 2026
Mean-difference activation steering allows for real-time emotional modulation in the Moshi full-duplex speech model without retraining. The study finds that while happy, angry, and surprise emotions share a linear steering direction, sadness is distinctly steerable in the residual stream.
HOW THIS AFFECTS YOU
●
builderYou can implement real-time emotional shifts in voice agents using simple vector additions to the residual stream.
●
designerThis enables more nuanced and reactive vocal personas in full-duplex conversational interfaces.