Results: Comparison with other methods & metrics

Below you'll find our results with an SVM model trained on FOR_rerec and FOR_2sec datasets and tested on the InTheWild dataset (available here). We can compare these results to those from the paper MLAAD: The Multi-Language Audio Anti-Spoofing Dataset.

Model Performance Comparison: All Models Trained on Identical Datasets

Model	Accuracy
Our SVM	68.9%
SLL W2V2	57.8%
Whisper DF	54.1%
RAWGAT-ST	49.8%

Performance Gain: Our SVM model shows an improvement of at least 11% compared to the best model (SLL W2V2) reported in the MLAAD paper.

The InTheWild dataset is available here.

---

detailed results

🔍 Overall Metrics

Metric	Value
Accuracy	0.6886
Precision	0.6834
Recall	0.6886
F1 Score	0.6801
ROC AUC	0.7484
Error Rate	0.3114

🧾 Classification Report

Class	Label	Precision	Recall	F1 Score	Support
0	Fake	0.7068	0.8142	0.7568	46,966
1	Real	0.6490	0.5042	0.5675	31,991