Paper deep dive
L-Proto: Language-Aware Episodic Prototypical Training for Multilingual Speaker Verification
Hyung-Seok Oh, Deok-Hyeon Cho, Seung-Bin Kim, Seong-Whan Lee
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 6/20/2026, 9:48:33 AM
Summary
L-Proto is a language-aware episodic prototypical training strategy designed to improve multilingual speaker verification. It addresses the issue of speaker-language entanglement, where embeddings capture language-specific cues instead of pure speaker identity. By constructing 'language-consistent episodes'—where each training episode contains speakers from only a single language—L-Proto reduces intra-episode linguistic variability and stabilizes prototype estimation. Experimental results on the TidyVoice Challenge benchmark demonstrate that L-Proto improves cross-lingual robustness and EER/minDCF performance across various backbone architectures like ResNet and ECAPA-TDNN compared to conventional fine-tuning and random episodic sampling.
Entities (6)
Relation Signals (4)
L-Proto → evaluatedon → TidyVoice Challenge
confidence 100% · Experiments on the TidyVoice Challenge benchmark demonstrate consistent performance improvements...
L-Proto → improves → Speaker Verification
confidence 100% · L-Proto yields the lowest EER... across all evaluated backbones.
SimAM-ResNet34 → isa → ResNet-based speaker embedding encoder
confidence 100% · We evaluate SimAM-ResNet34 and SimAM-ResNet100 [27], ResNet-based speaker embedding encoders...
L-Proto → addresses → Speaker-Language Entanglement
confidence 95% · L-Proto, a language-aware episodic prototypical training strategy... reduces language-driven variation during training and encourages embeddings to focus more directly on speaker identity.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Multilingual speaker verification remains challenging because language-dependent acoustic variability causes speaker identity to become entangled with linguistic characteristics, degrading generalization across languages. In multilingual training, embeddings often encode language cues with speaker identity, causing speakers to form language-specific clusters. We propose L-Proto, a language-aware episodic prototypical training strategy that constructs language-consistent episodes. By sampling speakers from a single language per episode, L-Proto reduces language-driven variation during training and encourages embeddings to focus more directly on speaker identity. Experiments on the TidyVoice Challenge benchmark demonstrate consistent performance improvements over conventional fine-tuning and random episodic sampling across multiple backbone architectures.
Tags
Links
- Source: https://arxiv.org/abs/2606.17416v1
- Canonical: https://arxiv.org/abs/2606.17416v1
Trouble viewing inline? Open PDF directly →
Full Text
28,009 characters extracted from source content.
Expand or collapse full text
Oh Cho Kim Lee L-Proto: Language-Aware Episodic Prototypical Training for Multilingual Speaker Verification Hyung-Seok Deok-Hyeon Seung-Bin Seong-Whan Department of Artificial Intelligence, Korea University, Seoul, Korea hs_oh@korea.ac.kr, sb-kim@korea.ac.kr, dh_cho@korea.ac.kr, sw.lee@korea.ac.kr Abstract Multilingual speaker verification remains challenging because language-dependent acoustic variability causes speaker identity to become entangled with linguistic characteristics, degrading generalization across languages. In multilingual training, embeddings often encode language cues with speaker identity, causing speakers to form language-specific clusters. We propose L-Proto, a language-aware episodic prototypical training strategy that constructs language-consistent episodes. By sampling speakers from a single language per episode, L-Proto reduces language-driven variation during training and encourages embeddings to focus more directly on speaker identity. Experiments on the TidyVoice Challenge benchmark demonstrate consistent performance improvements over conventional fine-tuning and random episodic sampling across multiple backbone architectures. keywords: speaker verification, multilingual speech, cross-lingual robustness, episodic training, prototypical learning 1 Introduction Speaker verification (SV) aims to determine whether two speech samples belong to the same speaker. Deep learning has significantly advanced this task by learning speaker-discriminative representations directly from speech signals [1, 2, 3], achieving strong performance on standard benchmarks [4, 5, 6, 7]. However, multilingual speaker verification remains challenging. Speaker identity is expressed through language-dependent phonetic and prosodic patterns, and training on multiple languages often causes speaker identity to become entangled with linguistic characteristics in the learned embeddings [8]. As a result, utterances from the same speaker may form language-specific subclusters, and verification performance often degrades when enrollment and test utterances are in different languages or when language distributions are imbalanced [9, 10, 11, 12]. To address cross-lingual variability, prior work has explored several complementary directions. Language-adversarial approaches introduce auxiliary classifiers with gradient reversal to encourage language-invariant embeddings [13, 14, 15, 16]. Other studies explicitly disentangle speaker and language factors through structured objectives or contrastive learning [17, 18]. Domain generalization frameworks further improve robustness to language shifts using meta-learning or distributionally robust optimization [19, 20, 21], while language-aware score calibration methods have also been investigated [22, 8]. In addition, large-scale multilingual pretraining has been shown to enhance cross-lingual robustness by exposing models to diverse linguistic conditions during representation learning [7, 23, 24]. Despite their differences, these approaches primarily operate at the representation level through global objectives applied across the full multilingual training set. In contrast, we focus on controlling linguistic variability at the task level by shaping the language composition of episodic training. Another line of research investigates metric-learning and episodic training strategies for SV. Prototypical networks perform distance-based classification by forming speaker prototypes within episodic tasks [25], and subsequent studies have explored improved support–query partitioning and optimization strategies for stable episodic learning [26]. Episodic supervision enhances local discrimination among sampled speakers. Most existing methods construct episodes by randomly sampling speakers from the training set [25, 26]. In multilingual settings, such episodes may contain speakers and utterances from multiple languages. When languages are mixed within an episode, embeddings of the same speaker may form language-dependent sub-clusters, reflecting speaker–language entanglement. This can distort prototype estimation and affect similarity comparisons within the episode, reducing the reliability of episodic supervision in multilingual settings. However, how to construct episodic tasks that remain reliable under multilingual speaker–language entanglement has received relatively less attention. These observations suggest that controlling linguistic variability at the task level can be beneficial for episodic learning in multilingual speaker verification. To this end, we propose L-Proto, a language-aware episodic prototypical training framework that constructs language-consistent episodes. Each episode contains speakers from a single language, reducing intra-episode linguistic variability and stabilizing prototype-based similarity learning. Different languages are sampled across training iterations, ensuring that the model remains exposed to multilingual data throughout training. Experimental results further suggest that L-Proto mitigates speaker–language sub-clustering, encouraging embeddings of the same speaker to remain more consistent across languages. The main contributions of this work are summarized as follows: • We show that language mixing within episodes can bias prototype estimation and destabilize similarity-based supervision in multilingual speaker verification. • We propose L-Proto, a language-aware episodic training strategy that constructs language-consistent episodes for multilingual speaker verification. • Experiments on TidyVoice demonstrate consistent gains across diverse backbones and cross-lingual scenarios, with code publicly available at https://github.com/hs-oh-prml/L-Proto/. Algorithm 1 Language-aware Episodic Sampling 1: Initialize buffers ℬ[ℓ][s]←∅B[ ][s]← 2: Initialize ready speaker sets ℛ[ℓ]←∅R[ ]← 3: for each incoming sample (x,s,ℓ)(x,s, ) do 4: x: speech sample, s: speaker ID, ℓ : language ID 5: if ℓ is invalid then 6: continue 7: end if 8: ℬ[ℓ][s]←ℬ[ℓ][s]∪xB[ ][s] [ ][s]∪\x\ 9: if |ℬ[ℓ][s]|≥K|B[ ][s]|≥ K then 10: ℛ[ℓ]←ℛ[ℓ]∪sR[ ] [ ]∪\s\ 11: end if 12: for each language ℓ′ with |ℛ[ℓ′]|≥P|R[ ]|≥ P do 13: Randomly sample P distinct speakers s1,…,sP\s_1,…,s_P\ ⊂ℛ[ℓ′] [ ] 14: Initialize episode ℰ←∅E← 15: for each selected speaker sis_i do 16: Randomly select K samples xi,1,…,xi,K⊂ℬ[ℓ′][si]\x_i,1,…,x_i,K\ [ ][s_i] 17: ℰ←ℰ∪(xi,k,si)k=1KE ∪\(x_i,k,s_i)\_k=1^K 18: ℬ[ℓ′][si]←ℬ[ℓ′][si]∖xi,1,…,xi,KB[ ][s_i] [ ][s_i] \x_i,1,…,x_i,K\ 19: if |ℬ[ℓ′][si]|<K|B[ ][s_i]|<K then 20: ℛ[ℓ′]←ℛ[ℓ′]∖siR[ ] [ ] \s_i\ 21: end if 22: end for 23: Output episode ℰE 24: end for 25: end for 2 Method We propose L-Proto, a language-aware episodic prototypical training framework for multilingual speaker verification. By constructing language-consistent episodes, it stabilizes prototype estimation during episodic learning. The framework consists of language-aware episode construction, streaming episode sampling, and episodic prototypical optimization. 2.1 Problem formulation Let the multilingual training set be defined as =(xi,yi,ℓi)i=1N,D=\(x_i,y_i, _i)\_i=1^N, where xix_i denotes an input speech sample, yiy_i the speaker label, and ℓi _i the language label. The speaker encoder fθ(⋅)f_θ(·) maps an input sample to an embedding vector i=fθ(xi)∈ℝd.z_i=f_θ(x_i) ^d. Standard training optimizes a global speaker classification objective over all samples in D, encouraging inter-speaker separability across languages. In contrast, episodic training introduces local discrimination among sampled speaker subsets. L-Proto modifies this sampling procedure by enforcing language consistency within each episode. 2.2 Language-aware episode construction To reduce cross-lingual variability within episodic training, we construct language-consistent episodes. For a given language ℓ , we sample P distinct speakers and K utterances per speaker to form an episode ℰℓ=(x(s,k),ys),E_ =\(x_(s,k),y_s)\, where s∈1,…,Ps∈\1,…,P\ indexes speakers and k∈1,…,Kk∈\1,…,K\ indexes utterances. All samples within ℰℓE_ share the same language label, so intra-episode variation is primarily determined by speaker identity rather than linguistic differences. This reduces cross-lingual interference during prototype estimation, allowing similarity comparisons to focus more on speaker identity. Although each episode contains a single language, episodes from different languages appear across training iterations, preserving exposure to multilingual data throughout training. This design reduces linguistic variability within each task while still allowing the encoder to observe diverse language conditions across tasks. 2.3 Streaming episode sampling To implement this episode construction strategy in practice, we employ a streaming sampling mechanism. Each buffer ℬ[ℓ][s]B[ ][s] stores utterances from speaker s in language ℓ . When a speaker accumulates at least K utterances, it becomes a ready speaker for that language. Once at least P ready speakers are available within the same language, an episode is formed by randomly selecting P speakers and K utterances per speaker. Algorithm 1 summarizes the sampling procedure. This strategy generates language-consistent episodes while preserving stochasticity over speakers and utterances. Episodes are formed on-the-fly during data loading without offline pre-grouping. 2.4 Episodic prototypical objective Given an episode ℰℓE_ , embeddings are computed as =fθ(x)z=f_θ(x). For each speaker s, a prototype is defined as the mean of its support embeddings: s=1|s|∑∈s,p_s= 1|S_s| _z _sz, (1) where sS_s denotes the support set of speaker s. For a query embedding q, cosine similarity to each prototype is computed as sim(,s)=⊤s|||s|.sim(q,p_s)= q p_s|q||p_s|. (2) The episodic loss is defined as a cross-entropy objective over similarity scores: ℒepi=−logexp(sim(,yq)/τ)∑s′=1Pexp(sim(,s′)/τ),L_epi=- (sim(q,p_y_q)/τ) _s =1^P (sim(q,p_s )/τ), (3) where yqy_q is the ground-truth speaker label of the query and τ denotes the temperature parameter used to scale cosine similarities. This objective enforces proximity between queries and their corresponding prototypes within each episode. Since episodes are language-consistent, similarity comparisons are less influenced by cross-lingual variability. Figure 1: t-SNE visualization of embeddings for a representative speaker subset under three training settings: (a) Pretrained, (b) Fine-tuning, and (c) L-Proto. Points are colored by speaker identity and edge color indicates language, showing that language-dependent sub-clusters emerge under conventional fine-tuning but become more consistent across languages with L-Proto. 2.5 Joint training objective The overall training objective combines global classification and episodic supervision: ℒ=ℒcls+λℒepi,L=L_cls+ _epi, (4) where ℒclsL_cls denotes the speaker classification loss, and λ controls the weight of episodic supervision. Model parameters θ are optimized end-to-end using this joint objective. Table 1: Performance on the TidyVoice Challenge development set under different language-mismatch conditions. Model Target / Non-Target All D/D D/S S/D S/S EER minDCF SimAM-ResNet34 1.56 4.81 0.95 2.81 2.88 0.85 w/ Fine-tuning 1.62 4.94 0.85 3.03 2.91 0.81 w/ L-Proto 0.83 2.24 0.55 1.69 1.38 0.63 SimAM-ResNet100 2.24 5.58 1.33 3.29 3.48 0.81 w/ Fine-tuning 1.17 4.70 0.51 2.39 2.63 0.77 w/ L-Proto 0.67 2.01 0.39 1.43 1.18 0.61 Table 2: Comparison across public speaker embedding backbones: pretrained, fine-tuned, and L-Proto models. Model Pretrain Fine-tuning L-Proto EER minDCF EER minDCF EER minDCF ResNet34 3.23 0.88 3.09 0.82 2.58 0.78 ResNet152 2.94 0.88 3.27 0.85 2.49 0.74 ResNet221 3.32 0.92 3.28 0.86 2.61 0.76 ResNet293 3.04 0.86 3.28 0.85 2.44 0.75 ECAPA512 4.15 0.96 4.52 0.92 3.73 0.95 ECAPA1024 4.70 1.00 4.95 0.96 3.69 0.92 CAM++ 4.06 0.93 3.90 0.92 2.50 0.76 3 Experiments 3.1 Experimental setup We evaluate SimAM-ResNet34 and SimAM-ResNet100 [27], ResNet-based speaker embedding encoders with simple attention modules, using the wespeaker toolkit [28]. The models are initialized from checkpoints pretrained on VoxBlink2 [7], fine-tuned on TidyVoiceX [11], and evaluated on the official TidyVoice Challenge development set. TidyVoiceX contains over 4,474 speakers, around 40 languages, and approximately 321k utterances from Mozilla Common Voice. The challenge evaluates similarity scores on same- and different-speaker trials under same- and cross-lingual conditions. We report EER and minDCF on the development set because official evaluation labels are not publicly available. Fine-tuning is performed for 6 epochs following the standard configuration used in the TidyVoice Challenge baseline setup. The learning rate is decayed from 5×10−55× 10^-5 to 1×10−51× 10^-5. Input features are 80-dim log Mel-filterbanks extracted from 16 kHz audio with 600-frame random cropping. For episodic training, each episode contains P speakers with two utterances per speaker. We set P=24P=24 and λ=100λ=100 for ResNet34, and P=8P=8 and λ=200λ=200 for ResNet100, with a temperature of 0.07. Hyperparameters were selected on the development set. Experiments are conducted on two NVIDIA RTX A6000 GPUs. All models are trained under the same data split and evaluation protocol for fair comparison. 3.2 Overall performance Table 1 summarizes performance under four trial language conditions. S and D denote same- and different-language pairs, respectively. Across both backbone architectures, L-Proto improves EER and minDCF over pretrained and fine-tuned baselines. The gains are most pronounced when trials involve cross-lingual target pairs (D/D and D/S), where cross-lingual variability directly affects similarity estimation. By constructing language-consistent episodes, L-Proto reduces language-induced variation in prototype-based discrimination while preserving multilingual coverage during training. L-Proto also improves matched-language conditions (S/D and S/S), indicating better speaker separability without degrading same-language performance. Table 3: Centroid cosine similarity analysis. SimintraSimintra measures cross-lingual centroid similarity within the same speaker, while SiminterSiminter measures same-language centroid similarity across different speakers. Δ denotes their separation margin. Method Simintra↑Sim_intra Siminter↓Sim_inter Δ↑ Pretrained 0.8027 0.2477 0.5550 Fine-tuning 0.7892 0.1048 0.6844 L-Proto 0.8536 0.0479 0.8057 3.3 Generalization across backbone architectures Table 2 evaluates L-Proto on a diverse set of publicly available pretrained speaker embedding models111https://github.com/areffarhadi/wespeaker/blob/master/docs/pretrained.md. They include VoxCeleb-pretrained ResNet variants [29], CAM++ [30], and ECAPA-TDNN variants [2]. For each model, we compare the original pretrained checkpoint, conventional fine-tuning on TidyVoiceX, and fine-tuning with L-Proto. Across all evaluated backbones, L-Proto yields the lowest EER. This suggests that language-aware episode construction is not tied to a particular architecture. For minDCF, L-Proto also yields lower values in most cases, indicating improved calibration of similarity scores under language mismatch. The magnitude of improvement varies across architectures, likely reflecting differences in model capacity and pretraining conditions. 3.4 Speaker–language disentanglement Figure 1 visualizes embedding distributions under different training strategies. Points are colored by speaker identity, while edge colors indicate language. In pretrained and fine-tuned models, utterances from the same speaker tend to form language-specific sub-clusters. In contrast, L-Proto produces more coherent clusters across languages, reflecting improved cross-lingual consistency. The red boxes illustrate that L-Proto merges language-specific clusters of the same speaker more effectively. To quantify this observation, we analyze cosine similarities between speaker–language centroids as summarized in Table 3. We compute SimintraSim_intra between language-specific centroids of the same speaker and SiminterSim_inter between different speakers within the same language. L-Proto increases intra-speaker cross-lingual similarity while reducing inter-speaker similarity, resulting in a larger separation margin. Table 4: Effect of episodic sampling and prototype supervision. Episode Prototype EER minDCF × × 3.48 0.81 ✓ × 2.40 0.74 × ✓ 2.63 0.77 ✓ ✓ 1.18 0.61 Table 5: Sensitivity to the episodic loss weight λ. Method λ EER minDCF SimAM-ResNet34 0 2.70 0.76 50 1.45 0.64 100 1.38 0.63 200 1.46 0.64 SimAM-ResNet100 0 2.17 0.71 100 1.25 0.62 200 1.18 0.61 300 1.21 0.62 Table 6: Effect of episode construction strategy. Method EER minDCF Random sampling 1.54 0.66 2 languages 1.35 0.65 4 languages 1.69 0.69 1 language 1.18 0.61 3.5 Ablation study In Table 4, the setting without episodic sampling and prototype supervision corresponds to conventional fine-tuning with only the global classification objective. Table 4 evaluates episodic sampling and prototype-based supervision. Each component improves the baseline, and their combination gives the largest gain. Table 5 examines the effect of the episodic loss weight λ. Using the episodic loss improves over λ=0λ=0 and remains stable across values. Table 6 analyzes the impact of episode language composition. Multi-language episodes do not improve consistently and can degrade as language diversity increases, suggesting that mixed-language episodes introduce nuisance variation into prototype estimation. Restricting each episode to a single language yields the best results, showing that L-Proto benefits from both episodic supervision and language-consistent episode construction. This trend is consistent with our motivation that reducing intra-episode linguistic variability stabilizes prototype-based supervision. 3.6 Language-wise improvement over pretrained models Figure 2 compares the improvement in EER over the pretrained model for each language in the development set. The upper panel shows the EER reduction achieved by fine-tuning and by the proposed L-Proto training, while the lower panel indicates the amount of available training data for each language. Both fine-tuning and L-Proto substantially improve over the pretrained model across languages. L-Proto generally yields larger gains than fine-tuning, suggesting better adaptation to multilingual conditions. The language-wise results show that L-Proto improves most languages, although the magnitude of improvement varies, indicating that its benefit is not uniform across language conditions. Figure 2: Language-wise improvement over the pretrained model. Top: EER reduction. Bottom: training data size. 4 Conclusion This paper introduced L-Proto, a language-aware episodic prototypical training strategy for multilingual speaker verification. By constructing single-language episodes, L-Proto reduces intra-episode linguistic variation and stabilizes prototype-based similarity learning. Experiments on the TidyVoice Challenge benchmark show consistent improvements over fine-tuning and random episodic sampling across multiple backbones. The method requires language labels and sufficient speaker diversity within each language, and introduces additional sampling overhead. Future work will explore adaptive episode construction for balancing linguistic diversity and speaker discrimination. 5 Use of Generative AI Disclosure Generative AI tools (OpenAI GPT-5.2) were used for writing assistance, including language editing and minor revisions of the manuscript. All technical content and conclusions were developed by the authors. 6 Acknowledgments This work was partly supported by Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No. RS-2019-I190079, Artificial Intelligence Graduate School Program (Korea University), IITP-2026-RS-2025-02304828, Artificial Intelligence Star Fellowship Support Program to nurture the best talents. References [1] L. Wan, Q. Wang, A. Papir, and I. L. Moreno, ``Generalized end-to-end loss for speaker verification,′ in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, p. 4879–4883. [2] B. Desplanques, J. Thienpondt, and K. Demuynck, ``ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,′ in Interspeech, 2020, p. 3830–3834. [3] J. weon Jung, Y. Kim, H.-S. Heo, B.-J. Lee, Y. Kwon, and J. S. Chung, ``Pushing the limits of raw waveform speaker recognition,′ in Interspeech, 2022, p. 2228–2232. [4] A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, ``Voxceleb: Large-scale speaker verification in the wild,′ Computer Speech & Language, vol. 60, p. 101027, 2020. [5] J. S. Chung, A. Nagrani, and A. Zisserman, ``VoxCeleb2: Deep speaker recognition,′ in Interspeech, 2018, p. 1086–1090. [6] Y. Lin, X. Qin, G. Zhao, M. Cheng, N. Jiang, H. Wu, and M. Li, ``Voxblink: A large scale speaker verification dataset on camera,′ in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, p. 10 271–10 275. [7] Y. Lin, M. Cheng, F. Zhang, Y. Gao, S. Zhang, and M. Li, ``VoxBlink2: A 100k+ speaker recognition corpus and the open-set speaker-identification benchmark,′ in Interspeech, 2024, p. 4263–4267. [8] J. Thienpondt, B. Desplanques, and K. Demuynck, ``Cross-Lingual Speaker Verification with Domain-Balanced Hard Prototype Mining and Language-Dependent Score Normalization,′ in Interspeech 2020, 2020, p. 756–760. [9] R. Auckenthaler, M. Carey, and J. Mason, ``Language dependency in text-independent speaker verification,′ in IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), vol. 1, 2001, p. 441–444 vol.1. [10] H. Mandalapu, T. M. Elbo, R. Ramachandra, and C. Busch, ``Cross-lingual speaker verification: Evaluation on x-vector method,′ in Intelligent Technologies and Applications, S. Yildirim Yayilgan, I. S. Bajwa, and F. Sanfilippo, Eds. Cham: Springer International Publishing, 2021, p. 215–226. [11] A. Farhadipour, J. Marquenie, S. Madikeri, and E. Chodroff, ``Tidyvoice: A curated multilingual dataset for speaker verification derived from common voice,′ arXiv preprint arXiv:2601.16358, 2026. [12] A. Vaheb, A. J. Choobbasti, S. H. E. M. Najafabadi, and S. Safavi, ``Investigating language variability on the performance of speaker verification systems,′ in Speech and Computer, A. Karpov, O. Jokisch, and R. Potapova, Eds. Cham: Springer International Publishing, 2018, p. 718–727. [13] J. Rohdin, T. Stafylakis, A. Silnova, H. Zeinali, L. Burget, and O. Plchot, ``Speaker verification using end-to-end adversarial language adaptation,′ in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019, p. 6006–6010. [14] W. Xia, J. Huang, and J. H. Hansen, ``Cross-lingual text-independent speaker verification using unsupervised adversarial discriminative domain adaptation,′ in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019, p. 5816–5820. [15] Z. Chen, S. Wang, and Y. Qian, ``Adversarial domain adaptation for speaker verification using partially shared network,′ in Interspeech, 2020, p. 3017–3021. [16] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. March, and V. Lempitsky, ``Domain-adversarial training of neural networks,′ Journal of Machine Learning Research, vol. 17, no. 59, p. 1–35, 2016. [17] A. Srinivas Menon, R. P. Gohil, K. Tripathi, and P. Wasnik, ``Laspa: Language agnostic speaker disentanglement with prefix-tuned cross-attention,′ in Interspeech, 2025, p. 3623–3627. [18] M. Olijslager, S. S. M. Ziabari, and A. M. M. Alsahag, ``Causally disentangled contrastive learning for multilingual speaker embeddings,′ arXiv preprint arXiv:2602.01363, 2026. [19] H. Zhang, L. Wang, K. A. Lee, M. Liu, J. Dang, and H. Meng, ``Meta-generalization for domain-invariant speaker verification,′ IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, p. 1024–1036, 2023. [20] J. Li, M.-W. Mak, J. Rohdin, K. A. Lee, and H. Hermansky, ``Bayesian Learning for Domain-Invariant Speaker Verification and Anti-Spoofing,′ in Interspeech 2025, 2025, p. 1123–1127. [21] S. Yang, D. Das, J. Cho, H. Park, and S. Yun, ``Domain Agnostic Few-shot Learning for Speaker Verification,′ in Interspeech, 2022, p. 595–599. [22] J. Thienpondt, B. Desplanques, and K. Demuynck, ``Tackling the score shift in cross-lingual speaker verification by exploiting language information,′ in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, p. 7187–7191. [23] A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y. Saraf, J. Pino, A. Baevski, A. Conneau, and M. Auli, ``XLS-R: Self-supervised cross-lingual speech representation learning at scale,′ in Interspeech, 2022, p. 2278–2282. [24] S. Chen, C. Wang, Z. Chen, Y. Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y. Qian, Y. Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, ``WavLM: Large-scale self-supervised pre-training for full stack speech processing,′ IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, p. 1505–1518, 2022. [25] T. Ko, Y. Chen, and Q. Li, ``Prototypical networks for small footprint text-independent speaker verification,′ in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, p. 6804–6808. [26] Y. Chen, W. Guo, and B. Gu, ``Improved meta-learning training for speaker verification,′ in Interspeech, 2021, p. 1049–1053. [27] X. Qin, N. Li, C. Weng, D. Su, and M. Li, ``Simple attention module based speaker verification with iterative noisy label detection,′ in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, p. 6722–6726. [28] S. Wang, Z. Chen, B. Han, H. Wang, C. Liang, B. Zhang, X. Xiang, W. Ding, J. Rohdin, A. Silnova et al., ``Advancing speaker embedding learning: Wespeaker toolkit for research and production,′ Speech Communication, vol. 162, p. 103104, 2024. [29] K. He, X. Zhang, S. Ren, and J. Sun, ``Deep residual learning for image recognition,′ in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. [30] H. Wang, S. Zheng, Y. Chen, L. Cheng, and Q. Chen, ``CAM++: A fast and efficient network for speaker verification using context-aware masking,′ in Interspeech, 2023, p. 5301–5305.