Collapse, Not Invariance: Diagnosing Auxiliary Objectives in Speech Anti-Spoofing

Ksenia Lysikova; Kirill Borodin; Maxim Maslov; Grach Mkrtchian

cs.SDSep 30, 2026

Collapse, Not Invariance: Diagnosing Auxiliary Objectives in Speech Anti-Spoofing

Authors: Ksenia Lysikova, Kirill Borodin, Maxim Maslov, Grach Mkrtchian

Organizations: Lab260, Yerevan, Armenia · BitmanagerAI, Dubai, UAE · MTUCI, Moscow, Russia

Abstract

Speech anti-spoofing countermeasures degrade when the generator, codec or channel changes, and a common remedy is an auxiliary objective that shapes the embedding space; whether it does is invisible to EER, a pure ranking metric. We compare seven such objectives with cross-entropy over 113 runs on five corpora, AASIST3 at three seeds plus four pre-trained detectors, and measure the embedding space of the 24 AASIST3 runs directly. Raw augmentation displacement makes cosine consistency look effective, but the gain is a smaller space, not a more stable one: normalised by the spread, no configuration consistently improves on cross-entropy. Every trained space is dominated by the single decision axis expected for two classes, whose training-set structure does not transfer, and four runs collapse to a near-constant output that displacement rewards and EER reports as poor accuracy. No auxiliary objective keeps an advantage over cross-entropy across architectures, corpora and seeds.

Figures & tables

Objective	s1	s2	s3	mean ${\pm}$ sd	s1	s2	s3	mean ${\pm}$ sd	s1	s2	s3	mean ${\pm}$ sd	s1	s2	s3	mean ${\pm}$ sd
	Displacement $A$				Spread $S$				Ratio $R$				Silhouette
CE	2.31	1.90	4.22	$2.81{\pm}1.24$	5.06	3.27	6.97	$5.10{\pm}1.85$	0.46	0.58	0.61	$0.55{\pm}0.08$	0.219	0.164	0.107	$0.163{\pm}0.056$
MSE	1.38	2.01	4.28	$2.56{\pm}1.53$	3.11	3.26	8.67	$5.01{\pm}3.16$	0.44	0.61	0.49	$0.52{\pm}0.09$	0.219	0.117	0.144	$0.160{\pm}0.053$
MAE	3.63	0.12	0.10	$1.28{\pm}2.03$	9.06	0.17	0.16	$3.13{\pm}5.14$	0.40	0.70	0.63	$0.58{\pm}0.16$	0.065	0.006	0.002	$0.024{\pm}0.035$
cosine	1.09	1.78	2.29	$1.72{\pm}0.60$	2.03	3.23	3.57	$2.94{\pm}0.81$	0.54	0.55	0.64	$0.58{\pm}0.06$	0.119	0.095	0.138	$0.117{\pm}0.022$
NT-Xent	1.38	2.70	3.47	$2.52{\pm}1.06$	1.85	5.93	8.73	$5.50{\pm}3.46$	0.75	0.46	0.40	$0.53{\pm}0.19$	0.130	0.214	0.061	$0.135{\pm}0.077$

Table 1: Diagnostics of the

24

AASIST3 runs on

600

ASVspoof 5 evaluation utterances, per seed and as mean

\pm

sd over seeds (collapsed runs included). Bold marks the four collapsed runs (Section 4.2 ):

S

and silhouette expose them,

A

ranks them the most invariant and

R

is unremarkable. Text differences are computed before rounding.

Figure 1: PCA (top) and t-SNE (bottom) of the $600$ ASVspoof 5 evaluation embeddings for MAE , smooth and DANN , each at a seed that trained and one that collapsed. PCA keeps one axis scale per column, so a trained run is a thin curve and a collapsed run a single point; t-SNE rescales every panel and turns both into similar-looking clusters.

Objective	ASV5	ASV21DF	ITW	VW	RuASD	5-corpus mean
CE	$\mathbf{0.288}{\pm}0.090$	$0.364{\pm}0.038$	$\mathbf{0.314}{\pm}0.022$	$\mathbf{0.293}{\pm}0.026$	$0.489{\pm}0.033$	$\mathbf{0.349}{\pm}0.002$
MSE	$0.391{\pm}0.099$	$0.410{\pm}0.110$	$0.412{\pm}0.180$	$0.365{\pm}0.178$	$0.510{\pm}0.143$	$0.418{\pm}0.140$
MAE	$0.490{\pm}0.075$	$0.408{\pm}0.149$	$0.454{\pm}0.060$	$0.404{\pm}0.115$	$\mathbf{0.465}{\pm}0.072$	$0.444{\pm}0.093$
cosine	$0.383{\pm}0.031$	$\mathbf{0.354}{\pm}0.008$	$0.368{\pm}0.032$	$0.309{\pm}0.077$	$0.481{\pm}0.037$	$0.379{\pm}0.023$
NT-Xent	$0.371{\pm}0.043$	$0.400{\pm}0.068$	$0.331{\pm}0.068$	$0.303{\pm}0.089$	$0.493{\pm}0.083$	$0.380{\pm}0.065$
smooth	$0.417{\pm}0.097$	$0.389{\pm}0.077$	$0.394{\pm}0.116$	$0.310{\pm}0.170$	$0.492{\pm}0.093$	$0.400{\pm}0.096$

Table 2: AASIST3 EER per corpus, mean

\pm

sd over seeds

1

–

3

(abbreviations in Section 2.2 ); the last column averages the five corpora per seed, then over seeds. Bold: lowest in the column (ties share it).

	DF-Arena			XLSR-SLS			Res2TCN			WhisperMFCC
Objective	full	LoRA	head	full	LoRA	head	full	LoRA †	head	full	LoRA	head
CE	0.454	0.488	0.570	0.362	0.492	0.565	0.349	–	0.341	0.444	0.505	0.520
MSE	0.375	0.564	0.567	0.500	0.498	0.568	0.374	–	0.333	0.541	0.513	0.527
MAE	0.417	0.567	0.565	0.372	0.487	0.562	0.363	–	0.331	0.508	0.506	0.493
cosine	0.383	0.566	0.566	0.387	0.517	0.566	0.376	–	0.343	0.528	0.511	0.475
NT-Xent	0.376	0.487	0.562	0.538	0.470	0.560	0.364	–	0.329	0.499	0.504	0.522

Table 3: Five-corpus mean EER of the fine-tuned detectors, one run each. Bold: lowest in the column (ties share it), an observed order, not a significant difference. Grey: collapsed encoders (Section 4.2 ), reported at chance.

Explore similar work

A Training-Efficient Transformer-Based Anti-Spoofing Network for Logical Access in ASVspoof 5
Jun 2, 2026Sidan Yin, Bo ZhaoSpoofing Automatic Speaker Verification
San Domenico School, San Anselmo, CA, United States · University of Washington, Seattle, WA, United States
Linguistic Bias Mitigation for Spoofing Detection via Gradient Reversal and A Variational Information Bottleneck
Jun 30, 2026Anh-Tuan Dao, Driss Matrouf, Mickael Rouvier +1Spoofing Linguistics
Laboratoire Informatique d’Avignon, Avignon Universite, France · EURECOM, Sophia Antipolis, France
When EER Hides Deployment Failure: Auditing Threshold Transfer and Unlabeled Score Calibration for Speech Deepfake Detectors
Jun 19, 2026Jingwen Zhou, Mingzhe WangEqual Error Rate In-Domain Accuracy
Xidian University, Xi’an, China

Jun 2, 2026, cs.SDJ/K move · Enter open · S save

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

Sidan Yin, Bo Zhao

San Domenico School, San Anselmo, CA, United States · University of Washington, Seattle, WA, United States

Spoofing Automatic Speaker Verification Face Forgery Detection Transformer Encoder Data Augmentation Access Control

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

Linguistic Bias Mitigation for Spoofing Detection via Gradient Reversal and A Variational Information Bottleneck

When EER Hides Deployment Failure: Auditing Threshold Transfer and Unlabeled Score Calibration for Speech Deepfake Detectors