Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Joint Optimization of ASV and CM tasks: BTUEF Team's Submission for WildSpoof Challenge

Oguzhan Kurnaz, Jagabandhu Mishra, Tomi Kinnunen, Cemal Hanilci

TL;DR

This work tackles robust spoofing-aware speaker verification by jointly optimizing ASV and CM components within a modular, non-linear fusion framework. The authors implement an end-to-end trainable SASV backend that can reuse public ASV and CM models, with a calibrated scoring scheme and an $a$-DCF–guided objective to align with evaluation metrics. Key findings show that a pretrained ReDimNet ASV backbone combined with a fine-tuned SSL-AASIST CM representation yields the best evaluation performance (eval a-DCF around 0.0515, final eval around 0.2163), outperforming official baselines. The results demonstrate the value of modular, condition-aware optimization for SASV in unconstrained spoofing scenarios, and highlight the benefit of exchanging ASV backbones for improved robustness.

Abstract

Spoofing-aware speaker verification (SASV) jointly addresses automatic speaker verification and spoofing countermeasures to improve robustness against adversarial attacks. In this paper, we investigate our recently proposed modular SASV framework that enables effective reuse of publicly available ASV and CM systems through non-linear fusion, explicitly modeling their interaction, and optimization with an operating-condition-dependent trainable a-DCF loss. The framework is evaluated using ECAPA-TDNN and ReDimNet as ASV embedding extractors and SSL-AASIST as the CM model, with experiments conducted both with and without fine-tuning on the WildSpoof SASV training data. Results show that the best performance is achieved by combining ReDimNet-based ASV embeddings with fine-tuned SSL-AASIST representations, yielding an a-DCF of 0.0515 on the progress evaluation set and 0.2163 on the final evaluation set.

Joint Optimization of ASV and CM tasks: BTUEF Team's Submission for WildSpoof Challenge

TL;DR

This work tackles robust spoofing-aware speaker verification by jointly optimizing ASV and CM components within a modular, non-linear fusion framework. The authors implement an end-to-end trainable SASV backend that can reuse public ASV and CM models, with a calibrated scoring scheme and an -DCF–guided objective to align with evaluation metrics. Key findings show that a pretrained ReDimNet ASV backbone combined with a fine-tuned SSL-AASIST CM representation yields the best evaluation performance (eval a-DCF around 0.0515, final eval around 0.2163), outperforming official baselines. The results demonstrate the value of modular, condition-aware optimization for SASV in unconstrained spoofing scenarios, and highlight the benefit of exchanging ASV backbones for improved robustness.

Abstract

Spoofing-aware speaker verification (SASV) jointly addresses automatic speaker verification and spoofing countermeasures to improve robustness against adversarial attacks. In this paper, we investigate our recently proposed modular SASV framework that enables effective reuse of publicly available ASV and CM systems through non-linear fusion, explicitly modeling their interaction, and optimization with an operating-condition-dependent trainable a-DCF loss. The framework is evaluated using ECAPA-TDNN and ReDimNet as ASV embedding extractors and SSL-AASIST as the CM model, with experiments conducted both with and without fine-tuning on the WildSpoof SASV training data. Results show that the best performance is achieved by combining ReDimNet-based ASV embeddings with fine-tuned SSL-AASIST representations, yielding an a-DCF of 0.0515 on the progress evaluation set and 0.2163 on the final evaluation set.
Paper Structure (5 sections, 1 equation, 1 figure, 1 table)

This paper contains 5 sections, 1 equation, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Proposed Method
  3. Experimental Setup
  4. Results
  5. Conclusion

Figures (1)

  • Figure 1: Proposed modular SASV architecture with trainable score-level integration. The pretrained ReDimNet-based ASV extractor is frozen, while only the AASIST component is trained in the SSL-AASIST branch with the SSL encoder frozen. ASV and CM embeddings are converted into calibrated scores, and all modules inside the red dashed block are trainable, producing the final SASV score via nonlinear fusion.