Duplex Dialogue Turn-Taking on Full-Duplex-Bench
0.01Synthetic TOR for Pause HandlingGPT-Realtime
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-Realtime2026.05 | 0.01 | 0.12 | 0 | 0.01 | 0.98 | 1 | 1.47 | 97 | 3.85 | 1.5 | — | |
| Listen-Write-Speak2026.05 | 0.01 | 0.01 | 0.53 | 0 | 1 | 0.97 | 0.48 | 99 | 4.02 | 0.65 | — | |
| DuplexCascadesystem backchannels=false2026.03 | 0.058 | 0.222 | 0.218 | 0.009 | 0.949 | 0.832 | 1.724 | 95.5 | 4.016 | 1.225 | 85.8 | |
| Gemini Livesource=directly obtained from the paper2026.03 | 0.255 | 0.31 | 0.091 | 0.012 | 0.896 | 0.655 | 1.301 | 89.1 | 3.376 | 1.183 | 77.8 | |
| Gemini Live 2.0 (deprecated)2026.05 | 0.26 | 0.31 | 0.09 | 0.01 | 0.9 | 0.66 | 1.3 | 89 | 3.38 | 1.18 | — | |
| DuplexCascade-βsystem backchannels=true2026.03 | 0.343 | 0.458 | 0.309 | 0.034 | 0.811 | 0.899 | 0.567 | 95 | 4.011 | 0.85 | 74.8 | |
| PersonaPlexsource=directly obtained from the paper2026.03 | 0.358 | 0.431 | 0.273 | 0.042 | 0.662 | 0.908 | 0.17 | 95 | 4.29 | 0.24 | 75.9 | |
| PersonaPlex2026.05 | 0.58 | 0.66 | 0.33 | 0.03 | 0.65 | 0.99 | 0.07 | 100 | 4.21 | 0.4 | — | |
| MiniCPM-Duplexsource=inferred from official checkpoints2026.03 | 0.613 | 0.731 | 0.582 | 0.003 | 0.96 | 0.916 | 0.61 | 100 | 1.14 | 0.447 | 59.8 | |
| Freeze-Omni2026.05 | 0.64 | 0.48 | 0.64 | 0 | 1 | 0.34 | 0.95 | 87 | 3.62 | 1.41 | — | |
| Freeze-Omnisource=directly obtained from the paper2026.03 | 0.642 | 0.481 | 0.636 | 0.001 | 0.997 | 0.336 | 0.953 | 86.7 | 3.615 | 1.409 | 48.9 | |
| dGSLM2026.05 | 0.93 | 0.94 | 0.69 | 0.02 | 0.93 | 0.98 | 0.35 | 92 | 0.2 | 2.53 | — | |
| dGSLMsource=directly obtained from the paper2026.03 | 0.934 | 0.935 | 0.691 | 0.015 | 0.934 | 0.975 | 0.352 | 91.7 | 0.201 | 2.531 | 46.6 | |
| Moshi2026.05 | 0.98 | 0.98 | 1 | 0 | 0.96 | 0.94 | 0.27 | 100 | 0.77 | 0.26 | — | |
| Moshisource=directly obtained from the paper2026.03 | 0.985 | 0.98 | 1 | 0.001 | 0.957 | 0.941 | 0.265 | 100 | 0.765 | 0.257 | 39.5 | |
| DoubaoEvaluation Mode=real-time mode, Latency (s)=1.822026.06 | — | — | — | — | — | — | — | — | — | — | 27.8 | |
| DuplexOmniEvaluation Mode=real-time mode, Latency (s)=0.5062026.06 | — | — | — | — | — | — | — | — | — | — | 72.6 | |
| Gemini-3.1-Flash-LiveEvaluation Mode=real-time mode, Latency (s)=2.572026.06 | — | — | — | — | — | — | — | — | — | — | 24.1 | |
| MiniCPM-o 4.5Evaluation Mode=real-time mode, Latency (s)=0.5022026.06 | — | — | — | — | — | — | — | — | — | — | 36.3 | |
| Qwen3-Omni-Realtime-FlashEvaluation Mode=real-time mode, Latency (s)=1.282026.06 | — | — | — | — | — | — | — | — | — | — | 25.2 | |
| Qwen3.5-Omni-Realtime-FlashEvaluation Mode=real-time mode, Latency (s)=1.252026.06 | — | — | — | — | — | — | — | — | — | — | 26.4 |