Beyond accuracy: user perception of differential privacy in speech emotion recognition

Abstract Speech emotion recognition (SER) is increasingly adopted in interactive systems, yet its deployment raises persistent concerns regarding privacy and user trust. Although differential privacy (DP) offers formal mathematical guarantees with quantifiable performance trade-offs, most existing SER implementations evaluate these trade-offs solely through technical metrics, leaving open the question of whether they align with users’ perceptions and expectations. This disconnect between technical privacy assurances and user acceptance motivates the need for privacy design approaches that foreground user understanding, control, and trust. Guided by a research-through-design approach, we first conducted a formative study ($N = 16$) to examine users’ privacy concerns and expectations surrounding SER technologies. Building on these insights, we designed a two-stage differentially private SER framework and instantiated it in an interactive online platform that integrates DP at both the model training and recognition stages with user-level privacy controls. Through an A/B study ($N = 32$), we investigated how varying privacy configurations influence perceived usefulness, trust, security, and overall acceptance. Our findings indicate that transparent and configurable privacy mediation can significantly enhance user trust and perceived security without materially compromising recognition performance. By conceptualizing privacy as a design material rather than a purely technical constraint, this work articulates user-centered principles for privacy mediation in emotion-aware AI systems and contributes to broader discussions on responsible interaction design, human–data relations, and the governance of affective technologies.