Full-duplex turn-taking on DuplexSLA-Bench (context-prefill)
96Accuracy (Normal)DuplexSLA
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| DuplexSLAContext prefill=true2026.05 | 96 | 0.27 | 93.33 | 0.27 | 99.33 | 0.4 | 98.33 | 0.32 | |
| gemini-3.1-flash-liveContext prefill=true2026.05 | 93.67 | 1.18 | 94.33 | 1.17 | 63.67 | 0.62 | 40 | — | |
| gpt-realtime-1.5Context prefill=true, VAD configuration=semantic-vad-high2026.05 | 91.33 | 1.67 | 90.33 | 1.68 | 79 | 0.68 | 0.33 | — | |
| gpt-realtime-1.5Context prefill=true, VAD configuration=server-vad-40ms2026.05 | 82.33 | 0.95 | 71 | 1.02 | 77 | 0.72 | 13 | — |