Voice AI's next bottleneck: turn-taking, not latency
机器之心 · wechat · 2026-11-15
A Machine Heart PRO deep-dive on voice AI's shift from latency to turn-taking. End-to-end delay is now 150-300ms, matching the 200ms human turn-taking gap, yet systems still interrupt or leave awkward silences because endpointing relies on silence thresholds alone. The piece surveys full-duplex approaches, Zhejiang University's L0-L3 architecture taxonomy, continuous prediction, and open problems in evaluation, controllability, and long-run stability.
More from Multimodal
- User asks Fable 5.2 to draw its self-portrait as SVG — it builds a universe out of letters — almostsweet · 2026-11-15
- AI cinematic Warcraft fan film starring Arthas resurfaces — tensorqt · 2026-11-15
- Dreamina launches Seedance 2.5 Flash: faster video generation from $0.0031/sec — aziz4ai · 2026-11-15
- Pirates of the Caribbean reimagined as a 1974 film, made with AI video generation — techhalla · 2026-11-15
- ComfyUI custom nodes automate animated typography faster than realtime — TomLikesRobots · 2026-11-15
- Free course maps the 4 ways generative media demos break in production — Al_Grigor · 2026-11-15