Full-duplex Speech Interaction Latency Analysis on Full-Duplex-Bench v1.5
0.68Stop Latency (Mean)Moshi (Fine-tuned)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Moshi (Fine-tuned)Scenario=Background Speech2026.03 | 0.68 | 0.48 | 0.73 | 0.49 | 44 | 192 | |
| Moshi (Fine-tuned)Scenario=User Backchannel2026.03 | 0.7 | 0.4 | 1.12 | 0.85 | 57 | 156 | |
| Moshi (Fine-tuned)Scenario=Talking to Other2026.03 | 0.82 | 0.65 | 0.84 | 0.69 | 47 | 188 | |
| Moshi (Fine-tuned)Scenario=User Interruption2026.03 | 0.89 | 0.64 | 0.66 | 0.55 | 110 | 383 | |
| Moshi (Base)Scenario=Background Speech2026.03 | 1.02 | 0.55 | 2.9 | 2.05 | 150 | 89 | |
| Moshi (Base)Scenario=Talking to Other2026.03 | 1.13 | 0.57 | 3.22 | 1.87 | 184 | 117 | |
| Moshi (Base)Scenario=User Backchannel2026.03 | 1.3 | 0.42 | 2.38 | 1.84 | 113 | 29 | |
| Moshi (Base)Scenario=User Interruption2026.03 | 1.3 | 0.72 | 1.99 | 2.24 | 391 | 237 |