Audio Spoof Detection: A Comprehensive Analysis of Multi-Model Approaches in SafeSpeak-2024

Overview

This project investigates advanced techniques for detecting audio spoofing using state-of-the-art self-supervised learning (SSL) and transformer architectures. It evaluates various models on ASVspoof benchmark datasets, demonstrating significant improvements in distinguishing between genuine and synthesized speech.

Models Evaluated

1. Wav2Vec 2.0
2. HuBERT
3. SSL Wav2Vec 2.0 with PSFAN Backend
4. Audio Spectral Transformer (AST)
5. Sound Event Detection Model (EfficientNet-B0)
6. WavLM Base

Key Data Augmentation Techniques

• Gaussian noise injection
• Signal-to-noise ratio modifications
• Dynamic gain variations
• Background noise injection

Loss Functions & Optimization

• Focal Loss, Cross-Entropy Loss, BCEWithLogits Loss
• Optimizer: AdamW with linear/cosine schedulers

Experimental Results

The Audio Spectral Transformer and SSL Wav2Vec outperformed other models with near-perfect precision and recall, demonstrating the power of transformer architectures in spoof detection.

Model	Public LB EER	Precision	Recall	F1-Score
Wav2Vec 2.0	0.46516	0.888	0.788	0.835
SSL Wav2Vec	0.02925	-	-	-
Audio Spectral Transformer	0.01384	0.999	0.999	0.999
HuBERT	8.11672	0.877	0.764	0.817
Pretrained Wav2Vec	0.77492	0.845	0.725	0.780
WavLM Base	1.87658	0.820	0.690	0.750

Links

• AST Spoofing Model (HuggingFace)
• SSL Wav2Vec Model (HuggingFace)
• GitHub Repository

Notebooks

• AST - 0.01384 EER on Public LB
• SSL - 0.02925 EER on Public LB