|
|
--- |
|
|
language: |
|
|
- en |
|
|
base_model: |
|
|
- microsoft/wavlm-base |
|
|
pipeline_tag: audio-classification |
|
|
datasets: |
|
|
- jungjee/asvspoof5 |
|
|
tags: |
|
|
- anti-spoofing |
|
|
- asvspoof5 |
|
|
--- |
|
|
# 🔎 Hybrid Pruning for Anti-Spoofing Results |
|
|
|
|
|
- **Input Feature**: Raw waveform (via SSL model) |
|
|
- **Frame Configuration**: 150 frames per segment, 20 ms frame shift |
|
|
- **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. |
|
|
- **Evaluation Metrics**: minDCF, EER (%) |
|
|
- **Evaluation Sets**: Dev / Eval |
|
|
- **Back-end**: Multi-Head Factorized Attentive Pooling (MHFA) |
|
|
|
|
|
--- |
|
|
|
|
|
# **Results on ASVspoof 5** |
|
|
|
|
|
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. |
|
|
|
|
|
| System | Dev minDCF | Dev EER (%) | Eval minDCF | Eval EER (%) | |
|
|
| :--- | :--- | :--- | :--- | :--- | |
|
|
| Rank 3 (ID:T27, Fusion) | - | - | 0.0937 | 3.42 | |
|
|
| **HP (ours, Single system)** | 0.0395 | 1.547 | **0.1028** | **3.758** | |
|
|
| Rank 4 (ID:T23, Fusion) | - | - | 0.1124 | 4.16 | |
|
|
| Rank 9 (ID:T23, Best single system) | - | - | 0.1499 | 5.56 | |
|
|
|
|
|
|
