Trait-State Disentanglement

Identity leakage was much lower in MP-IB state embeddings than in baseline representations. The paper treats the information-capacity bound as empirical support rather than a complete disentanglement guarantee. The article interprets residual Top-1 identity signal as more related to demographics than individual identity. State monitoring requires representations that suppress anatomy, habitual prosody, and speaker identity while preserving agitation-linked speech dynamics. The article cautions that demographic leakage may still matter in privacy-sensitive clinical deployments.