A Latency-Cover Architecture for Enhancing Perceived Responsiveness in AI VTubers

Accumulated latency from language generation, speech synthesis, and avatar rendering can introduce noticeable silence in live AI VTuber broadcasts. This study reframes this problem in terms of perceived responsiveness and proposes a dual-track latency-cover architecture. FastTrack immediately plays a pre-synthesized preliminary reaction grounded in user emotion and dialogue act, while SlowTrack generates the context-aware main response. In a repeated-measures evaluation with 32 participants, Character Appeal differed significantly across conditions (Friedman p = .009), and the Emotion Only condition was rated higher than the SlowTrack Only condition (Holm-adjusted p = .043); however, most other measures showed no significant condition effects. The proposed architecture shifts the focus from computational optimization alone to perceptual interaction design and provides a framework for evaluating the contextual appropriateness of preliminary reactions.