A Comparison of SSL-Based Feature Extractors and Back-End Classifiers for Spoofing Detection: A Multi-Corpus Training and Cross-Linguistic Analysis

Authors: Anh-Tuan Dao, Driss Matrouf, Mickael Rouvier, Nicholas Evans

Published: 2026-06-07 15:20:38+00:00

AI Summary

This paper conducts a comprehensive benchmark of four self-supervised learning (SSL) feature extractors paired with four back-end classifiers for voice spoofing detection. Through multi-corpus training and cross-linguistic analysis, the study exposes domain biases within datasets and highlights the importance of language-specific adaptation. The findings emphasize the critical need for domain-aware and language-specific approaches to enhance spoofing detection robustness.

Abstract

Voice biometric systems face growing threats from spoofing attacks, yet the evaluation of detection models remains inconsistent across datasets. To investigate these unpredictable fluctuations, we conduct a comprehensive benchmark of four self-supervised learning feature extractors paired with four back-end classifiers. We compare the hierarchical local feature extraction of ResNet with the global sequence and relational modeling of attention and graph-based back-ends. Through multi-corpus training across three scenarios and six evaluation datasets, our empirical analysis yields two critical findings. First, we expose a domain bias within the ASVspoof 5 dataset, showing that naive data scaling actively degrades performance. Second, our cross-linguistic analysis reveals that fine-tuning with just 8 hours of target-language data enhances detection robustness. Together, these findings emphasize the critical need for domain-aware and language-specific adaptation in spoofing detection.

Key findings

The proposed ResNet back-end classifier consistently achieved the lowest average Equal Error Rate (EER), demonstrating superior performance. XLSR was identified as the most effective SSL feature extractor, owing to its extensive and diverse pre-training data. Multi-corpus training exposed significant domain biases within datasets like ASVspoof 5, showing that naive data scaling can degrade performance, while fine-tuning with even modest amounts of target-language data (e.g., 8 hours) markedly improved cross-linguistic detection robustness.

Approach

The authors benchmark four SSL-based feature extractors (Wav2vec2, HuBERT, WavLM, XLSR) with four back-end classifiers (AASIST, Conformer, MHFA, and a proposed ResNet architecture). They train models across three multi-corpus scenarios and evaluate them on six diverse datasets, including cross-linguistic data, to analyze performance fluctuations and generalization capabilities.

Datasets

ASVspoof 5 training set, MLAAD-v3, ASVspoof19, VCTK (for training); ASVspoof21 LA Hidden, ASVspoof21 DF Hidden, ASVspoof 5 evaluation set, Fake-Or-Real, HABLA (Spanish), CFAD (Chinese) (for evaluation); MUSAN corpus, real room impulse response (RIR) database (for data augmentation).

Model(s)

Wav2vec2-Large-LV60K, HuBERT-Large, WavLM-Large, Wav2vec2-XLSR-300m (XLSR) as SSL feature extractors; AASIST, Conformer, MHFA, and a proposed 34-layer ResNet as back-end classifiers.

Author countries

France

← Previous