Update README.md
Browse files
README.md
CHANGED
|
@@ -9,4 +9,26 @@ datasets:
|
|
| 9 |
tags:
|
| 10 |
- anti-spoofing
|
| 11 |
- asvspoof5
|
| 12 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
tags:
|
| 10 |
- anti-spoofing
|
| 11 |
- asvspoof5
|
| 12 |
+
---
|
| 13 |
+
## 🔎 Hybrid Pruning for Anti-Spoofing Results
|
| 14 |
+
|
| 15 |
+
- **Input Feature**: Raw waveform (via SSL model)
|
| 16 |
+
- **Frame Configuration**: 150 frames per segment, 20 ms frame shift
|
| 17 |
+
- **Training Strategy**: Jointly optimizing for task performance and model sparsity in a single stage. A warm-up schedule is used where the sparsity target linearly increases from 0 to the final value over the first 5 epochs.
|
| 18 |
+
- **Evaluation Metrics**: minDCF, EER (%)
|
| 19 |
+
- **Evaluation Sets**: Dev / Eval
|
| 20 |
+
- **Back-end**: Multi-Head Factorized Attentive Pooling (MHFA)
|
| 21 |
+
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
### **Results on ASVspoof 5**
|
| 25 |
+
|
| 26 |
+
The following table compares the performance of our proposed **Hybrid Pruning (HP) single system** against other top-performing systems from the official ASVspoof 5 Challenge leaderboard.
|
| 27 |
+
|
| 28 |
+
| System | Dev minDCF | Dev EER (%) | Eval minDCF | Eval EER (%) |
|
| 29 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 30 |
+
| Rank 3 (ID:T27, Fusion) | - | - | 0.0937 | 3.42 |
|
| 31 |
+
| **HP (ours, Single system)** | 0.0395 | 1.547 | **0.1028** | **3.758** |
|
| 32 |
+
| Rank 4 (ID:T23, Fusion) | - | - | 0.1124 | 4.16 |
|
| 33 |
+
| Rank 9 (ID:T23, Best single system) | - | - | 0.1499 | 5.56 |
|
| 34 |
+
|