Super Star: Towards Streaming Real-time Interactive Agents for Digital Humans
The paper targets the missing real-time layer for digital humans: gestures generated live from streaming speech, without future audio.
Super Star pairs a streaming speech response module with a causal gesture generator that uses current audio and motion history. The authors say this avoids the latency and lookahead assumptions of offline co-speech gesture systems. They also describe synthetic training data for virtual companion dialogues, shaped by topic and emotion, plus a feedback loop for adapting to user preferences. Experiments report better latency-quality trade-offs, stronger speech-motion sync, and higher user preference than baselines. HF Daily Papers' note
Super Star pairs a streaming speech response module with a causal gesture generator that uses current audio and motion history. The authors say this avoids the latency and lookahead assumptions of offline co-speech gesture systems. They also describe synthetic training data for virtual companion dialogues, shaped by topic and emotion, plus a feedback loop for adapting to user preferences. Experiments report better latency-quality trade-offs, stronger speech-motion sync, and higher user preference than baselines. HF Daily Papers' note
score 4