Language impact assessment and domain shift mitigation in audio spoofing detection

Life Cycle Reliability and Safety Engineering · Springer Nature · 2026-09-23 · DOI: 10.1007/s10772-026-10294-5

Аннотация

This paper presents a systematic study on Thai speech spoofing detection, addressing unique challenges posed by tonal languages through deep feature analysis with robustness against various attacks. We systematically evaluate feature extraction methods (Constant Q Cepstral Coefficients (CQCC), Linear Frequency Cepstral Coefficients (LFCC), WavLM with 50% Principal Component Analysis (PCA)), model architectures (Light Convolutional Neural Network (LCNN), Res2Net, Custom Convolutional Neural Network (CNN)), data augmentation strategies, and unsupervised domain adaptation (UDA) techniques to identify optimal configurations. Experiments used the ASVspoof 2019 Logical Access (LA) dataset and custom-generated Thai/English synthetic samples with attack characteristics aligned to ASVspoof specifications. Results reveal LFCC features consistently outperform traditional CQCC and self-supervised WavLM representations across all architectures. The baseline Custom CNN with Convolutional Block Attention Module (CBAM) attention achieved 4.00% Equal Error Rate (EER) on the official ASVspoof 2019 LA evaluation set. UDA with gradient reversal layer (GRL) and a 10% unlabeled target proportion improved the mean EER to 1.86% ± 0.2167% over five independent runs. The lowest single run across the 5%–30% settings was 1.21% EER (20% unlabeled target). This UDA setting is transductive, because unlabeled evaluation samples are used during adaptation. Attack-wise results show that the gain is largest on mismatched evaluation attack A17, with a smaller gain on A18. A controlled synthetic Thai/English check (550 samples per language) is reported separately. Under matched ASVspoof-inspired attack simulations, language differences were small, and UDA models reached 100% accuracy on that limited set. These synthetic results are limited to matched generation conditions and are not interpreted as real-world language independence. Overall, the main finding is that GRL-based UDA reduces domain shift on ASVspoof 2019 LA relative to the same Custom CNN + CBAM baseline.

Читать у издателя →

Похожие статьи