Update README.md
Browse files
    	
        README.md
    CHANGED
    
    | @@ -3,6 +3,25 @@ license: cc-by-4.0 | |
| 3 | 
             
            ---
         | 
| 4 |  | 
| 5 | 
             
            # SEA Fake Speech Dataset
         | 
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
|  | |
| 6 |  | 
| 7 | 
             
            ## Dataset Summary
         | 
| 8 | 
             
            This dataset contains multilingual speech data for deepfake detection,
         | 
|  | |
| 3 | 
             
            ---
         | 
| 4 |  | 
| 5 | 
             
            # SEA Fake Speech Dataset
         | 
| 6 | 
            +
            The rapid growth of the digital economy in South-East Asia (SEA)
         | 
| 7 | 
            +
            has amplified the risks of audio deepfakes—yet current datasets
         | 
| 8 | 
            +
            cover SEA languages only sparsely, leaving models poorly equipped
         | 
| 9 | 
            +
            to handle this critical region. This omission is critical: detection
         | 
| 10 | 
            +
            models trained on high-resource languages collapse when applied
         | 
| 11 | 
            +
            to SEA, due to mismatches in synthesis quality, language-specific
         | 
| 12 | 
            +
            characteristics, and data scarcity. To close this gap, we present
         | 
| 13 | 
            +
            SEA-Spoof, the first large-scale ADD dataset especially for SEA
         | 
| 14 | 
            +
            languages. SEA-Spoof spans 300+ hours of paired real and spoof
         | 
| 15 | 
            +
            speech across Tamil, Hindi, Thai, Indonesian, Malay, and Viet-
         | 
| 16 | 
            +
            namese. Spoof samples are generated from a diverse mix of state-
         | 
| 17 | 
            +
            of-the-art open-source and commercial systems, capturing wide
         | 
| 18 | 
            +
            variability in style and fidelity. Benchmarking state-of-the-art detec-
         | 
| 19 | 
            +
            tion models reveals severe cross-lingual degradation, but fine-tuning
         | 
| 20 | 
            +
            on SEA-Spoof dramatically restores performance across languages
         | 
| 21 | 
            +
            and synthesis sources. These results highlight the urgent need for
         | 
| 22 | 
            +
            SEA-focused research and establish SEA-Spoof as a foundation
         | 
| 23 | 
            +
            for developing robust, cross-lingual, and fraud-resilient detection
         | 
| 24 | 
            +
            systems.
         | 
| 25 |  | 
| 26 | 
             
            ## Dataset Summary
         | 
| 27 | 
             
            This dataset contains multilingual speech data for deepfake detection,
         | 
