Found first: a primary source the press has not covered yet.
A new paper introduces Speculative Listening, a training-free framework that uses idle decoder compute during audio listen windows in full-duplex speech models to pre-draft the next response chunk, then verifies the draft against actual arriving audio. Published as arXiv:2611.06982, the work finds that 95.3% of drafted chunks are unaffected by the real audio, making the approach nearly lossless in practice.
What the source says
Yutong Bian and colleagues at Chinese University of Hong Kong, Shenzhen describe the core insight: full-duplex speech models interleave listen and speak windows in fixed-time chunks, leaving the decoder idle while audio streams in. Speculative Listening fills that idle window by decoding under a silence hypothesis, then verifying and correcting the draft once real audio arrives. When correction is needed, 88% of the time it affects only the first output token, keeping rollback cost low. Speak chunks accelerate 2.6 to 2.8x; tool-call chunks accelerate by up to 23.4x. Response-start latency falls 73% on average to a range of 74 to 241 ms, approaching the roughly 200 ms gap in human conversation, and maximum real-time lag drops 94.4%. The paper introduces TickTock-Bench, a per-chunk latency benchmark evaluated across eight tasks.
Why it matters
Every deployed full-duplex voice agent burns idle decoder cycles while audio arrives. Speculative Listening converts that dead time into useful pre-computation with no changes to model weights or training procedure. The 95.3% verification pass rate means the speculative work is almost always kept, so the overhead of the rare correction is minimal. Tool-call chunks, often the highest-latency steps in voice agent pipelines, see the largest gains. The method requires no retraining and can be applied to any chunk-interleaved full-duplex system already in production.