Speculative Listening speeds full-duplex voice AI 2.6 to 2.8x

Full-duplex speech models, which allow simultaneous listening and speaking, typically interleave listen and speak windows in fixed-time chunks. During listen windows, the decoder sits idle while audio streams in, wasting compute cycles. Speculative Listening exploits this by decoding under a silence hypothesis during the idle window, then verifying and correcting the draft once real audio arrives. The paper reports that when correction is needed, 88% of the time it affects only the first output token, keeping rollback cost low. This approach contrasts with methods that require model changes or additional training. The introduction of TickTock-Bench, a per-chunk latency benchmark evaluated across eight tasks, may provide a standardized way to measure progress in this area. The work was published as arXiv:2611.06982 by Yutong Bian and colleagues at Chinese University of Hong Kong, Shenzhen.