Table 1
Performances of the proposed approaches on the DiCOVA Track-1 database. The performances on the validation sets are averaged over the five folds. SD = standard deviation.
| Validation |
Test |
|||
|---|---|---|---|---|
| (%) | Sp (SD) | AUC (SD) | Sp | AUC |
| Baseline [5] | – | 68.81 | – | 69.91 |
| Hand-crafted-feature-based representations | ||||
| Log Mel | 33.16 (8.49) | 60.67 (6.15) | 27.60 | 67.41 |
| MFCC | 30.88 (7.63) | 61.16 (3.65) | 23.44 | 55.63 |
| ComParE | 33.58 (8.08) | 65.50 (1.96) | 40.10 | 66.90 |
| Deep image-from-audio-based representations | ||||
| VGG11 | 32.23 (2.73) | 63.09 (2.55) | 42.71 | 65.69 |
| ResNet34 | 20.41 (3.83) | 51.69 (3.80) | 38.02 | 58.75 |
| Deep audio-based representations | ||||
| CNN14_16k | 42.28 (5.49) | 67.82 (3.29) | 44.27 | 67.77 |
| ResNet38 | 39.27 (14.81) | 68.36 (4.54) | 45.31 | 65.66 |
| Ensemble learning | ||||
| Feature-max | 00.00 (0.00) | 63.73 (3.94) | 28.12 | 65.56 |
| Feature-avg | 35.54 (8.59) | 68.51 (2.09) | 61.46 | 73.72 |
| Feature-attention | 44.56 (6.29) | 70.56 (3.01) | 59.38 | 77.96 |
| Decision-max | 31.40 (8.20) | 66.66 (2.25) | 38.02 | 66.44 |
| Decision-avg | 35.03 (9.67) | 68.57 (2.98) | 55.73 | 73.25 |
| Decision-attention | 39.17 (9.41) | 68.21 (3.92) | 59.38 | 77.36 |
Bold values: The feature-level attention (specificity: 59.38%, AUC: 77.96%) outperforms the other two feature-level fusions (i.e. feature-max and feature-average (avg)), and the decision-level attention (specificity: 59.38%, AUC: 77.36%) performs the best among the three decision-level fusions.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.