AA Speech to Speech - Conversational Dynamics (Full Duplex Bench): leaderboard
Artificial Analysis's Conversational Dynamics score for speech-to-speech systems: AA's weighted average of pause handling, turn-taking, interruption handling and backchannel handling, measured with subsets of Full Duplex Bench v1 and v1.5. It scores how a voice system manages the flow of a live conversation, not whether its answers are correct (that is the Big Bench Audio and Tau3-Voice boards). AA's 0-1 value times 100, higher is better.
Metric: Score (0-100), AA's weighted Full Duplex Bench v1/v1.5 subset: pause handling, turn-taking, interruption, backchannel. Source: artificialanalysis.ai. Saturation forecast: Rough model projection: around 2026. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen Audio 3.0 Realtime Plus | 98.41 | |
| 2 | GPT-Live-1 (Sol, low) | 97.27 | |
| 3 | Grok Voice 3.0 | 96.97 | |
| 4 | Qwen Audio 3.0 Realtime Flash | 96.95 | |
| 5 | Gemini 3.8 Live | 96.11 | |
| 6 | GPT-Realtime-2 (Minimal) | 96.06 | |
| 7 | Eleven Realtime v1 Mini | 95.82 | |
| 8 | GPT Realtime 2.1 (High) | 95.71 | |
| 9 | GPT-Realtime-1.5 | 95.7 | |
| 10 | GPT Realtime Mini (Oct '25) | 95.68 |
Interactive version: theaggregate.ai/benchmark?slug=aa-speech-to-speech-conversational-dynamics-full-duplex-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.