Counterfactual check stops speech models from talking during silence

A new inference-time method aims to stop full-duplex speech models from unexpectedly talking during prolonged user silence. Full-duplex systems such as Moshi and PersonaPlex can listen and speak at the same time, but sometimes initiate speech when no person has said anything.

Under digital-zero input, Moshi began speaking in 12 of 40 five-minute continuations and PersonaPlex in 11 of 40. At each observed onset, the probability of speech jumped by more than nine orders of magnitude within one 80-millisecond frame, suggesting an abrupt response to the model’s own preceding nonspeech output rather than gradual random sampling.

The proposed filter asks a counterfactual question: would the next-token distribution look similar if the preceding user input were muted? If so, it suppresses the onset as unrelated to genuine speech. No retraining is required.

In held-out tests with realistic microphone noise, the method blocked 13 of 13 false starts for Moshi and nine of nine for PersonaPlex while retaining 40 of 40 genuine responses for each model. Its 95th-percentile decision time was below 61 milliseconds, inside the frame budget. These are small experimental samples from a new arXiv preprint, so broader acoustic and conversational testing remains necessary.