Tag
This paper analyzes spurious onsets in full-duplex speech LLMs like Moshi and PersonaPlex and introduces an inference-time mitigation method based on causal analysis to suppress them without retraining.