Full-duplex voice models still fall short on humanlike zero-overlap turn-taking
alexisgallagher · x · 2026-09-13
The author notes that full-duplex audio models (like the one in the ChatGPT app) handle backchanneling more organically, but still don't achieve humanlike zero-overlap turn-taking, and none have shown convincing multi-speaker behavior.
Citing a recent benchmark paper: today's best turn predictors have poor recall (missing many genuine turn ends) and high false positives (mistaking pauses for turn ends). Worse, in real human conversation the turn gap is often negative — overlaps mean there is no gap at all — a fundamental challenge for current systems.
Related event: Study: voice AI turn-taking still far from human-level(3 posts)→
More from Models
- Leaked GPT-6 Sol Output Impresses, But OpenAI Reportedly Has Stronger Bell Internally — VraserX · 2026-09-14
- Chinese LLMs have caught up a lot, argues Kevin Bass — the US should widen its AI lead, not slow down — kevinnbass · 2026-09-14
- ARC-AGI-3's non-standard scoring under fire as GPT-6 hits ~100% — peterwildeford · 2026-09-14
- eigenrobot compares sol vs astra: tighter, colder reasoning, 'net less fun' but a strength — eigenrobot · 2026-09-14
- Users Report GPT5.6 Sol Has Gotten Noticeably Dumber Lately — VoidStateKate · 2026-09-14
- OpenAI's New Model Solved Navier-Stokes in 88 Hours — Then a Credit Fight Broke Out — Don't Worry About the Vase (Zvi) · 2026-09-13