Study: voice AI turn-taking still far from human-level

Human conversation flows with near-zero gaps or overlaps, yet a new benchmark shows even the best turn-prediction models miss many turn ends and produce false alarms. Full-duplex models like gpt-realtime feel more natural in backchannels but still fall short.

2026-09-13 ~ 2026-09-13 · 3 related posts

• Human conversation runs with near-zero gaps and overlaps — a bar no voice AI meets — literalbanana · 2026-09-13
• Turn prediction benchmark: best systems have poor recall and high false positives — alexisgallagher · 2026-09-13
• Full-duplex voice models still fall short on humanlike zero-overlap turn-taking — alexisgallagher · 2026-09-13