Full-duplex turn-taking on DuplexSLA-Bench (no-context-prefill average of 2 scenarios)
96.5Accuracy (%)gpt-realtime-1.5
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| gpt-realtime-1.5VAD setting=semantic-vad-high2026.05 | 96.5 | 1.57 | |
| DuplexSLA2026.05 | 94.34 | 0.3 | |
| gemini-3.1-flash-live2026.05 | 93.17 | 1.17 | |
| gpt-realtime-1.5VAD setting=server-vad-40ms2026.05 | 85.5 | 0.83 | |
| MiniCPM-o2026.05 | 82 | 0.61 | |
| PersonaPlex2026.05 | 22.34 | 0.47 | |
| Freeze-Omni2026.05 | 10.67 | 0.36 |