Hallucination detection on VOXPOPULI k=10% (test)
1.23Hallucination Rate (%)Attention Score
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Attention ScoreBackbone=Qwen-2-Audio2026.04 | 1.23 | 96 | 8 | 15 | 6 | 4 | 30 | |
| PerplexityBackbone=Qwen-2-Audio2026.04 | 3.18 | 97 | 50 | 51 | 49 | 49 | 43 | |
| Mean EntropyBackbone=Qwen-2-Audio2026.04 | 3.33 | 97 | 50 | 50 | 50 | 49 | 43 | |
| RAUQ EntropyBackbone=Qwen-2-Audio2026.04 | 3.42 | 96 | 46 | 45 | 47 | 47 | 46 | |
| Logistic RegressionBackbone=Qwen-2-Audio, Feature Set=AUDIORATIO, Number of Features=10242026.04 | 3.57 | 97 | 56 | 54 | 58 | 56 | 49 | |
| Logistic RegressionBackbone=Qwen-2-Audio, Feature Set=Combined, Number of Features=40962026.04 | 3.62 | 97 | 55 | 52 | 57 | 58 | 51 | |
| Logistic RegressionBackbone=Qwen-2-Audio, Feature Set=Top 75, Number of Features=3002026.04 | 3.87 | 97 | 52 | 49 | 57 | 58 | 49 | |
| RandomBackbone=Qwen-2-Audio2026.04 | 50.27 | 50 | 6 | 3 | 50 | — | — |