A Training-Efficient Transformer-Based Anti-Spoofing Network for Logical Access in ASVspoof 5

Authors: Sidan Yin, Bo Zhao

Published: 2026-06-02 00:38:12+00:00

Comment: 11 pages, 2 figures

AI Summary

This paper introduces TFPARN, a Transformer-based focal-pairwise attentive ranking network for training-efficient anti-spoofing in the ASVspoof 5 Track 1 closed condition. TFPARN extracts log-Mel features, uses a Transformer encoder, applies attention pooling, and is trained with a combination of focal classification loss and pairwise ranking loss, along with RawBoost augmentation. It achieves state-of-the-art detection accuracy while significantly improving computational efficiency compared to baselines.

Abstract

Synthetic and manipulated speech can reduce the reliability of automatic speaker verification systems, so anti-spoofing methods need to be both accurate and efficient in training and inference. This paper focuses on the ASVspoof 5 Track 1 closed condition, where standard cross-entropy training may not give enough attention to hard trials and is not directly aligned with ranking- and threshold-based evaluation metrics. We propose TFPARN, a Transformer-based focal-pairwise attentive ranking network. The system extracts log-Mel features from speech, uses a Transformer encoder to model frame-level information, applies attention pooling to obtain utterance-level representations, and is trained with a combination of focal classification loss and pairwise ranking loss. RawBoost augmentation is used during training, and test-time augmentation is applied during evaluation to improve robustness. Compared with re-implemented AASIST and RawNet2 baselines under the same protocol, TFPARN achieves the best results, with a minDCF of 0.2430 and an EER of 12.52%. Ablation experiments further show that the pairwise loss, focal loss, and attention pooling all improve performance. TFPARN also uses the lowest inference memory among the compared systems, at 1.4 GB, runs at about 0.79 ms per utterance, and reaches its best checkpoint in less training time than AASIST. These results show that TFPARN provides a good balance between detection accuracy and computational cost for logical access anti-spoofing.

Key findings

TFPARN achieves the best detection results with a minDCF of 0.2430 and an EER of 12.52%, outperforming AASIST and RawNet2. It also demonstrates superior computational efficiency, using the lowest inference memory (1.4 GB), running at about 0.79 ms per utterance, and reaching its best checkpoint in less training time than AASIST, indicating a better balance between accuracy and cost.

Approach

The proposed TFPARN system extracts log-Mel features from speech, which are then processed by a Transformer encoder to model frame-level information. Attention pooling aggregates these into utterance-level representations, which are fed into a classification head. The network is trained with a combined objective of focal classification loss (to emphasize hard samples) and pairwise ranking loss (to align with ranking-sensitive metrics).

Datasets

ASVspoof 5 Track 1 dataset, built from the English subset of the Multilingual LibriSpeech (MLS) corpus.

Model(s)

TFPARN (Transformer-based Focal-Pairwise Attentive Ranking Network) with a Transformer encoder and attention pooling. Compared against AASIST and RawNet2 baselines.

Author countries

UNKNOWN

← Previous