Integrated Spoofing-Robust Automatic Speaker Verification via a Three-Class Formulation and LLR

Kai Tan; Lin Zhang; Ruiteng Zhang; Johan Rohdin; Leibny Paola García-Perera; Zexin Cai; Sanjeev Khudanpur; Matthew Wiesner

eess.ASMar 14, 2026

Integrated Spoofing-Robust Automatic Speaker Verification via a Three-Class Formulation and LLR

Authors: Kai Tan, Lin Zhang, Ruiteng Zhang, Johan Rohdin, Leibny Paola García-Perera, Zexin Cai, Sanjeev Khudanpur, Matthew Wiesner,

Organizations: HLTCOE & 2 CLSP, Johns Hopkins University, USA · Tianjin University, China

Abstract

Spoofing-robust automatic speaker verification (SASV) aims to integrate automatic speaker verification (ASV) and countermeasure (CM). A popular solution is fusion of independent ASV and CM scores. To better modeling SASV, some frameworks integrate ASV and CM within a single network. However, these solutions are typically bi-encoder based, offer limited interpretability, and cannot be readily adapted to new evaluation parameters without retraining. Based on this, we propose a unified end-to-end framework via a three-class formulation that enables log-likelihood ratio (LLR) inference from class logits for a more interpretable decision pipeline. Experiments show comparable performance to existing methods on ASVSpoof5 and better results on SpoofCeleb. The visualization and analysis also prove that the three-class reformulation provides more interpretability.

Figures & tables

Fig. 1 : Comparison of SASV paradigms. (a) Fusion-based SASV typically trains separate ASV and CM using $\mathcal{L}_{\mathrm{spk}}$ and $\mathcal{L}_{\mathrm{cm}}$ , and may include a fusion loss $\mathcal{L}_{\mathrm{fusion}}$ in either score level (shown here) or embedding level. (b) Existing bi-encoder-based integrated SASV systems use a unified SASV objective $\mathcal{L}_{\mathrm{sasv}}$ , often together with ASV-related or contrastive objectives, and independently encode enrollment and test utterances before scoring. (c) The proposed 3T2-SASV adopts a cross-encoder architecture with a three-class formulation and converts the class logits into an SASV LLR for inference. Unlike the double- or single-score formulations in (a) and (b), the proposed method preserves explicit class-level evidence for the three SASV hypotheses. Gray boxes denote non-trainable modules.

Fig. 2 : Overview of the proposed 3T2-SASV framework. Given enrollment and test utterances, the embedding extractor first produces utterance-level representations, which are then aggregated to obtain a fused representation $\mathbf{F}_{t}$ . The three-class classifier predicts target, nontarget, and spoof logits, which are converted into an SASV log-likelihood ratio $\mathrm{llr}_{\mathrm{SASV}}$ for the final accept/reject decision.

	Emb. Extractor	Method	min a-DCF
ASVspoof 5	MFA-Conformer	ASV5-B04 [ 6 ]	0.5882
	MFA-Conformer	SA-SASV [ 17 ]	0.6025
	MFA-Conformer	Proposed 3T2-SASV	0.6029
	SKA-TDNN	ASV5-B04	0.5788
	SKA-TDNN	SA-SASV	0.5830
	SKA-TDNN	Proposed 3T2-SASV	0.5961

TABLE I : Comparison with existing methods (min a-DCF on evaluation sets).

Fig. 3 : Comparison of score distributions between the ASV5-B04 baseline and the proposed 3T2-SASV on SpoofCeleb.

Fig. 4 : Min a-DCF across different operating spoof priors for the fixed-prior and prior-adjusted LLRs.

Emb. Extractor	Dev.	Eval.
wav2vec 2.0 Base	0.2545	0.4821
Wav2vec 2.0 Large	0.2293	0.4510
MFA-Conformer	0.2126	0.5973

TABLE II : min a-DCF of wav2vec 2.0 and MFA-Conformer as embedding extractors for the proposed 3T2-SASV.

Num of Ref Utt	min a-DCF
1	0.2126
2	0.2337
3	0.2586
4	0.2366

TABLE III : The Effect of Different Number of Ref Utterances

Utt Length	min a-DCF
2.0	0.2126
3.0	0.2589
5.0	0.2624

TABLE IV : The Effect of Different Utterance Length

Number of Attention Heads	min a-DCF
1	0.2906
2	0.2984
4	0.2126
8	0.2787

TABLE V : The effect of different number of attention heads on the ASVspoof5 development set.

Strategy	min a-DCF
Random sampling	0.2481
Hard-pair	0.2126

TABLE VI : Comparison of sampling strategies on the development set of ASVspoof 5.

Fig. 5 : Comparison of three enrollment-test aggregation strategies in the proposed 3T2-SASV framework.

Aggregation Strategy	min a-DCF
Embedding Concatenation	0.4071
Waveform Concatenation	0.3297
Cross-attention	0.2126

TABLE VII : Effect of different enrollment-test aggregation strategies on the ASVspoof5 development set.

Scoring strategy	min a-DCF
Prior-adjusted LLR	0.5741
Calibrated LLR	0.5861

TABLE VIII : Effect of LLR calibration on the ASVspoof 5 evaluation set.

Fig. 6 : Class-conditional score distribution of the Integrated B04 method and the proposed method on ASVspoof 5.

Fig. 7 : Attack-wise score distribution of the spoof class in the proposed method on ASVspoof 5.

Speaker Verification

Speaker verification determines whether two speech recordings come from the same person, enabling voice authentication and identity analysis.

See all 62 papers

Large Audio Language Models for Spoofing-Aware Speaker Verification
Jul 16, 2026Sofya Savelyeva, Mariia Perunova, Evgeny Kushnir +3Audio-Language Model Evaluation Audio Deepfake Detection
Applied AI Institute · MIRAI · AXXX, HSE +3
LISE : Listenable Interpretable Speaker Embeddings
Jun 19, 2026Xiaoliang Wu, Chongxin Gan, Ke Liu +2Audio Representation Learning Speaker Verification
University of Southampton, United Kingdom · The Hong Kong Polytechnic University, Hong Kong SAR, China · University of Edinburgh, United Kingdom
Unified Shared Encoder in Spoof-Aware Speaker Verification with Hybrid Wavelet Prompt Tuning
Oct 1, 2026Aref Farhadipour, Srikanth Madikeri, Teodora Vukovic +2Speaker Verification ASVspoof
Department of Computational Linguistics, University of Zurich, Switzerland · Idiap Research Institute, Switzerland · Brno University of Technology, Czech Republic

Jul 16, 2026J/K move · Enter open · S save

Large Audio Language Models for Spoofing-Aware Speaker Verification

Sofya Savelyeva, Mariia Perunova, Evgeny Kushnir, Artem Dvirniak+2

Applied AI Institute · MIRAI · AXXX, HSE+3

Audio-Language Model Evaluation Audio Deepfake Detection Speaker Verification

Open paper AskPDF arXiv

Recent advances in text-to-speech and voice cloning make high-quality spoofing inexpensive and scalable, threatening voice authentication systems, especially automatic speaker verification (ASV). Existing defenses mainly address this threat through binary countermeasures (CMs) for deepfake detection or spoofing-aware speaker verification (SASV), where current systems are dominated by modular ASV-CM fusion and cascaded pipelines. Although large audio language models (LALMs) have shown promise on related audio tasks, including CM and ASV, their use for SASV remains unexplored, despite their capacity to produce natural-language rationales for auditing and robustness beyond discriminative predictions. This work systematically evaluates LALMs for SASV against conventional pipelines under zero-shot prompting, supervised adaptation, reasoning-oriented training, and reinforcement-learning-based optimization. Our results show that pretrained LALMs are near chance in the zero-shot setting, confirming that they are not natively suited to SASV, but that task-specific adaptation closes this gap. We further find that competitive SASV performance can be achieved through several distinct routes. These findings position LALMs as a promising and auditable foundation for unified SASV, while clarifying where conventional cascade systems still lead.

Abstract

Speaker Verification

Explore similar work

Large Audio Language Models for Spoofing-Aware Speaker Verification

LISE : Listenable Interpretable Speaker Embeddings

Unified Shared Encoder in Spoof-Aware Speaker Verification with Hybrid Wavelet Prompt Tuning