LASPA: Language Agnostic Speaker Disentanglement with Prefix-Tuned Cross-Attention

Aditya Srinivas Menon; Kumud Tripathi; Pankaj Wasnik; Raj Prakash Gohil

REVIEW 5 major objections 6 minor 40 references

LASPA: Language Agnostic Speaker Disentanglement with Prefix-Tuned Cross-Attention

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read LASPA is a training strategy that strips language-dependent information from speaker embeddings and improves equal error rate across monolingual and multilingual test sets.

desk verdict LASPA is a plausible and useful training strategy for language-agnostic speaker embeddings, but its central disentanglement claim rests on unvalidated pseudo-label diversity within VoxCeleb2 and an under-specified cross-lingual test protocol. read the letter →

arxiv 2506.02083 v1 pith:F2DLSJOO submitted 2025-06-02 cs.SD cs.AIcs.LGcs.MM

Aditya Srinivas Menon , Raj Prakash Gohil , Kumud Tripathi , Pankaj Wasnik This is my paper

classification cs.SDcs.AIcs.LGcs.MM

keywords speakerrecognitionlanguage-agnosticembeddingsdisentangledrepresentationlearningprefix-tuningcross-attentionmultilingualverificationdiarizationequalerrorrate

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Speaker recognition systems can mistake language-induced changes in a voice for a different speaker, so the same person speaking two languages may be scored as two people. LASPA is a training strategy that adds a language encoder, two prefix-tuned cross-attention modules, and a reconstruction decoder to push language-specific information out of the speaker embedding. The paper reports that this lowers equal error rate (the operating point where false-accept and false-reject rates cross) on monolingual and multilingual test sets, including the out-of-distribution NISP-B set, and raises the cosine similarity of the same speaker across languages. At inference the model is unchanged: only the speaker encoder is used, so the improvement comes from training rather than from a larger deployment-time model.

What carries the argument

The load-bearing mechanism is the pair of prefix-tuned cross-attention modules. In the speaker-to-language tuner, the query comes from the speaker embedding while keys and values come from the language embedding with learned prefix vectors appended; in the language-to-speaker tuner the roles are swapped. The attention outputs are fused back into each embedding, and the Mean Absolute Pearson Correlation loss drives the two embeddings toward statistical independence. Prefix tuning keeps almost all backbone parameters frozen, and the paper states the prefix-tuners make up only 1.16% of total model parameters, so the disentanglement is carried by a small set of new parameters on top of existing speaker and language encoders.

What would settle it

Take bilingual speakers with ground-truth language labels, train LASPA, and measure two things: equal error rate on cross-language speaker trials and accuracy of a language classifier trained on the resulting speaker embeddings. The central claim fails if the language classifier stays far above chance while equal error rate shows no improvement over the plain backbone.

Watch

Extended reading notes

Core claim

The paper claims that a joint-loss training loop can disentangle language from speaker identity without adversarial training. During training, a speaker encoder and a language encoder read the same mel-spectrogram; two prefix-tuned cross-attention modules let each embedding attend to the other with learned prefix vectors; and a decoder reconstructs the mel-spectrogram from the fused embeddings. Four losses are summed with equal weight: AAM-Softmax for speaker classification, negative log-likelihood for language classification under pseudo-labels, Mean Absolute Pearson Correlation to decorrelate the speaker and language embeddings, and mean squared error for reconstruction. After training, the language branch is discarded. The paper's evidence is that LASPA lowers equal error rate and minimum detection cost relative to the same backbone trained without it on VoxCeleb1, VoxCeleb1-B, VoxSRC 2020/2021, and NISP-B, lowers spoken-language-recognition accuracy read from speaker embeddings, and raises cross-language cosine similarity of the same speaker.

Load-bearing premise

The method relies on machine-generated guesses about which language is being spoken in the training audio, and it assumes those guesses are accurate and varied enough to teach the model what to remove; if the training set is close to monolingual or the guesses are wrong, the loss may strip away speaker information instead of language information.

Editorial extensions

If this is right

On the stronger ReDimNet backbone, LASPA lowers equal error rate from 2.15% to 2.10% on the VoxSRC 2021 validation set and from 1.66% to 1.62% on VoxCeleb1-B.
On the out-of-distribution NISP-B multilingual set, LASPA ReDimNet reaches 10.22% equal error rate versus 11.90% for the baseline, indicating transfer to languages not seen in training.
LASPA reduces spoken-language-recognition accuracy on VoxCeleb1-B from 81.2% to 78.5% and raises same-speaker cross-language cosine similarity from 0.612 to 0.641 for ReDimNet, consistent with less language information in the speaker embedding.
The ablation on ResNet-S shows that removing the prefix-tuners raises VoxCeleb1-B equal error rate from 5.88% to 6.44%, and removing the language branch entirely raises it to 7.00%.
Because inference uses only the speaker encoder, the disentanglement is a training-time intervention that leaves deployment compute unchanged.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because LASPA is training-only and leaves the speaker encoder architecture untouched, it is a candidate plug-in for any mel-spectrogram speaker encoder, not just the four backbones tested here.
Pseudo-label supervision is harder to obtain for code-switched utterances or closely related languages; a natural extension is to test LASPA with ground-truth language labels or self-supervised language clusters instead of spoken-language-recognition guesses.
The reported gains quantify removal of language information, but the method does not separate accent from language; in applications where accent is a legitimate speaker cue, aggressive decorrelation might suppress useful identity information.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is a clean empirical combination of known pieces: prefix-tuned cross-attention borrowed from NLP/multimodal work applied to speaker-language disentanglement, trained with MAPC, AAM softmax, MSE, and NLL losses. That specific configuration is new, and the inference encoder is unchanged, which is a practical plus. The gains are real on the reported numbers: consistent EER reductions across VoxCeleb1-B, VoxSRC, and the new NISP-B set, with the largest jumps on the weaker backbones and modest gains on ReDimNet. The ablation showing prefix-tuners add value beyond just adding a language encoder is genuinely useful, and the SLR accuracy drop on VoxCeleb1-B is directionally consistent with the stated goal.

The soft spots are real but mostly validation gaps rather than demonstrated failures. No error bars, no code or data, and the GRL baselines are only partially described, so I cannot tell if the comparison is apples to apples. The bigger concern is the one you flagged: the language pseudo-labels come from an SLR model applied to VoxCeleb2 dev, and the paper reports no statistics on the resulting label distribution. VoxCeleb2 is English-dominant enough that if the pseudo-labels collapse to English or a few noisy classes, the NLL loss cannot teach the language encoder anything language-specific, and the MAPC term is just decorrelating the speaker embedding from a near-constant auxiliary signal. In that case the EER gains would reflect generic regularization, not language disentanglement, and the "unseen languages" claim for NISP-B would be unsupported. The NISP-B protocol is described in one sentence, so there is no way to check the trials are balanced and cross-lingual as claimed.

That said, the paper is not circular: the losses target disentanglement directly, and the test sets are held out. The ablation helps, and the SLR drop is consistent with the mechanism. I would not reject the idea; I would ask for evidence that the pseudo-labels actually contain diverse languages before believing the mechanism.

Who gets value: anyone working on multilingual speaker verification or code-switched diarization who wants a cheap training-time trick. It deserves a serious referee, provided the authors are pressed to release the label distribution, error bars, and a fuller NISP-B specification. My recommendation: send it to peer review, but expect heavy revision before the language-disentanglement claim is supportable.

Referee Report

5 major / 6 minor

Summary. The paper proposes LASPA, a multi-task training strategy for language-agnostic speaker recognition. During training, a speaker encoder and a language encoder are trained jointly with two prefix-tuned cross-attention modules and a decoder that reconstructs the mel-spectrogram. The training objective combines an AAM-softmax speaker loss, an NLL language-classification loss using pseudo-labels from a VoxLingua107-trained SLR model, an MSE reconstruction loss, and a Mean Absolute Pearson's Correlation (MAPC) loss intended to decorrelate the speaker and language embeddings. At inference, only the speaker encoder is used. Experiments with ResNet-S, ResNet-L, ECAPA, and ReDimNet backbones report improved EER/minDCF on VoxCeleb1, VoxCeleb1-B, VoxSRC 2020/2021, and NISP-B, reduced SLR accuracy on VoxCeleb1-B, higher cosine similarity for same-speaker cross-lingual trials, and lower DER on DISPLACE. An ablation on ResNet-S shows that removing the prefix-tuners or the language-related components degrades EER.

Significance. If the results hold, LASPA would be a practically useful training-time method: it leaves the inference encoder unchanged, adds only a small parameter overhead via prefix tuners, and reports consistent gains across several backbones and benchmarks, including a cross-lingual evaluation dataset. The paper's strengths are the breadth of the evaluation, the directionally consistent SLR reduction reported in Table 2, and the inclusion of an ablation isolating the contribution of the language branch and prefix tuners. However, several load-bearing validation gaps remain: the diversity of the pseudo-labels used to supervise the language encoder is not established, the 'unseen languages' protocol for NISP-B is not specified, no uncertainty quantification is provided for the small EER differences, and the MAPC loss and GRL baselines are under-specified. These gaps currently prevent the paper from fully supporting the claim that the gains are due to language disentanglement rather than to a generic auxiliary/regularization effect.

major comments (5)

[Section 4.2 (Datasets)] The pseudo-label diversity on VoxCeleb2 is unvalidated. The language encoder is trained with the NLL and MAPC losses using pseudo-labels from an SLR model trained on VoxLingua107, but the paper does not report how many languages are represented in the VoxCeleb2 dev set, their frequency distribution, or the confidence/agreement of the pseudo-labels. VoxCeleb2 is commonly treated as an English-dominant corpus; if the pseudo-labels collapse to a small number of languages or are very noisy, the NLL term cannot teach a meaningful language embedding, and the MAPC term would simply decorrelate E_spk from a near-constant or noisy auxiliary signal. In that case the EER gains in Table 1 and Table 3 and the SLR drop in Table 2 could reflect a generic auxiliary-task/regularization effect rather than language disentanglement. Please report the pseudo-label distribution on VoxCeleb2 dev, including per-language counts and any available metadata agreement, and ideally include an ablation with curated or gold language labels if such labels are available.
[Section 4.2 and Table 1 (NISP-B protocol)] The claim that LASPA generalizes to 'unseen languages' is not supported by the described protocol. The NISP-B dataset is described in a single sentence: it is created from NISP, contains intra-speaker cross-lingual and inter-speaker monolingual trials, and is said to be balanced across English, Hindi, Kannada, Malayalam, Tamil, and Telugu. The paper does not specify how 'unseen' is defined, whether these six languages appear in the VoxLingua107-labeled training pseudo-labels, how cross-lingual trials are sampled to prevent speaker or language leakage, how many trials are used, or how the cosine scoring is performed. Without a precise evaluation protocol, the NISP-B results cannot be interpreted as evidence of generalization to unseen languages. Please provide a complete description of the NISP-B trial construction, language separation, and scoring protocol.
[Tables 1 and 3 (Uncertainty quantification)] No error bars, confidence intervals, or multiple-seed results are reported. Several of the claimed improvements are very small, for example ReDimNet EER 2.15 to 2.10 on VoxSRC 2021 validation, 1.66 to 1.62 on VoxCeleb1-B, and 0.37 to 0.31 on VoxCeleb1 clean. Without a measure of variance, the statement that LASPA 'consistently outperforms' baselines is not statistically grounded. Please report standard deviations or bootstrap confidence intervals over at least three training runs, or otherwise demonstrate that the differences are not within the expected run-to-run variation.
[Equation (2) and Section 3 (MAPC loss definition)] The MAPC loss, which is central to the disentanglement claim, is never defined. The paper names 'Mean Absolute Pearson's Correlation' and cites reference [9], but no formula or precise description is given. It is unclear whether the correlation is computed per dimension of E_spk and E_lng, averaged over the batch, or applied to pooled embeddings, and whether gradients flow to both encoders. Without this definition the main training objective is not reproducible. Please provide the exact mathematical form of L_MAPC and state how it is computed over the mini-batch.
[Section 5.1 and Table 1 (GRL baselines)] The GRL baselines are insufficiently described. The paper states that GRL-based approaches are prone to hyperparameters and unstable training, but it does not report the GRL implementation details: where the gradient reversal layer is inserted, the reversal weight schedule, the language-classification loss used, or how the GRL models are trained relative to the same recipe (1200 epochs, batch size 400, etc.). Moreover, GRL results are absent on NISP-B, which is the key cross-lingual out-of-distribution benchmark. Without these details the comparison against the main competing disentanglement approach is not fully reproducible, and the claimed advantage over GRL on unseen languages is not measured.

minor comments (6)

[Equation (1)] The notation in Equation (1) is unclear: 'ppJKp' should presumably be the square root of the key dimension J_{K_p}; please define all symbols (Q, K_p, V_p, P_k, P_v) explicitly and fix the typesetting.
[Table 1 caption] The caption says 'The two version of LASPA with and without prefix-tuner is compared,' but Table 1 only reports one LASPA row per backbone; the no-prefix-tuner variant appears only in the Table 3 ablation. Please correct the caption to match the content.
[Table 2] The 'LASPA' row aggregates results across backbones, while the other rows are per model; please report SLR accuracy and cosine similarity per backbone for LASPA, and describe how SLR accuracy is measured (model, training set, and evaluation split) so that the 78.5% figure is interpretable.
[Section 4.1] The statement that each configuration is trained for 1200 epochs needs qualification: please specify whether early stopping is used, which validation set (if any) selects the checkpoint, and whether the reported numbers are from the final or best checkpoint.
[References] Reference [2] appears to have a corrupted author string ('S. M. E. M. A. Conkie'); please verify the citation metadata.
[Figure 1 and Section 3.3] Figure 1 contains 'PK', 'PV', and 'FCFEATURE EXTRACTOR' notations that are not defined in the text; please harmonize the figure with the notation used in Equation (1) and the surrounding text.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: empirical training method, held-out evaluation, and non-load-bearing self-citations.

full rationale

LASPA is an empirical training method; the final loss (Eq. 2) directly defines the optimization objective, and the reported EER, SLR, cosine-similarity, and DER values are external measurements on test sets (VoxCeleb1, VoxCeleb1-B, VoxSRC 2020/2021, NISP-B, DISPLACE) that are disjoint from the VoxCeleb2 dev training set. The equal loss weights are a stated hyperparameter choice, not a parameter fitted to the target and then renamed as a prediction. The language pseudo-label supervision and VoxLingua107 initialization create a dependency between the auxiliary 'language' definition and the SLR model used to generate labels, but the central speaker-verification claim is not equivalent to that dependency by construction: the NISP-B and VoxCeleb1-B improvements are held-out outcomes, and Table 2's SLR drop is an independent probe. The self-citations ([22], [31]) are not load-bearing: [22] is cited only to motivate prefix-tuned cross-modal attention, and [31] is a prior DISPLACE entry cited alongside the external challenge baseline [30]. No uniqueness theorem or ansatz is imported from the authors' own prior work to force the architecture. The lack of reported pseudo-label distribution on VoxCeleb2 is an unvalidated assumption and a correctness risk, but it is not a circular reduction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim depends on standard losses and dataset assumptions rather than on new theoretical entities. The most fragile premises are the reliability of SLR pseudo-labels and the assumption that MAPC decorrelation preserves speaker identity. No new particle, force, or conserved quantity is introduced.

free parameters (4)

Loss weights for L_MSE, L_AS, L_MAPC, L_NLL = 1.0 each (equal weights)
Section 4: 'We empirically found that using equal weights led to optimal performance across evaluation metrics.' This is a hyperparameter choice without sensitivity analysis.
Prefix length and number of attention heads in prefix-tuners
Not specified in the paper; these affect the capacity of the cross-attention fusion and are chosen implicitly by the authors.
Embedding dimension J for speaker and language embeddings
Section 3.1 and 3.2: 'transforms the mel-spectrogram X_i into a J dimensional vector.' J is never given.
Training hyperparameters (learning rate, weight decay, epochs, batch size) = lr=1e-3, wd=2e-5, epochs=1200, batch=400
Section 4.1 gives these values, but they are not justified as optimal; they are standard choices.

assumptions (5)

domain assumption Pseudo-language labels from an SLR model trained on VoxLingua107 are accurate for VoxCeleb2 dev set.
Section 4.2. The speaker encoder is trained with these labels via NLL loss; if labels are noisy or non-diverse, the language encoder learns little.
domain assumption Minimizing MAPC between speaker and language embeddings reduces linguistic information in the speaker embedding without losing speaker identity.
Section 4, loss L_MAPC. This is the core mechanism of disentanglement, but the paper does not prove it; it only measures SLR accuracy post-hoc.
standard math AAM Softmax on speaker labels yields speaker-discriminative embeddings.
Section 4, loss L_AS. Standard loss from face recognition [11], widely used in speaker verification.
domain assumption Mel-spectrogram reconstruction via LSTM decoder forces the combined embeddings to retain sufficient information to regenerate the input.
Section 3.4 and loss L_MSE. The reconstruction objective is assumed to preserve speaker-relevant acoustic detail.
ad hoc to paper Equal weighting of the four losses is near-optimal.
Section 4: 'We empirically found that using equal weights led to optimal performance.' This is a hyperparameter choice without sensitivity analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LASPA: Language Agnostic Speaker Disentanglement with Prefix-Tuned Cross-Attention." pith.science (2026). https://pith.science/paper/F2DLSJOO

@misc{pith2026250602083,
  author       = {Pith},
  title        = {Pith review of: LASPA: Language Agnostic Speaker Disentanglement with Prefix-Tuned Cross-Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F2DLSJOO}},
  note         = {Machine review of arXiv:2506.02083}
}

read the original abstract

Speaker recognition models face challenges in multi-lingual settings due to the entanglement of linguistic information within speaker embeddings. The overlap between vocal traits such as accent, vocal anatomy, and a language's phonetic structure complicates separating linguistic and speaker information. Disentangling these components can significantly improve speaker recognition accuracy. To this end, we propose a novel disentanglement learning strategy that integrates joint learning through prefix-tuned cross-attention. This approach is particularly effective when speakers switch between languages. Experimental results show the model generalizes across monolingual and multi-lingual settings, including unseen languages. Notably, the proposed model improves the equal error rate across multiple datasets, highlighting its ability to separate language information from speaker embeddings and enhance recognition in diverse linguistic conditions.

Figures

Figures reproduced from arXiv: 2506.02083 by the authors.

**Figure 1.** Block diagram of the proposed architecture. Speaker and Language embeddings are extracted using the speaker and language encoder respectively and then after it is fused through prefix-tuners, and then passed to the decoder for mel-spectogram reconstruction. resentation. Recently, there has been growing interest in speaker recognition tasks in multi-lingual scenarios. In this direction, [1] addressed the challenge … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 35 canonical work pages

[9]

Our method effectively separates speaker iden- tity from linguistic information, addressing key challenges in multi-lingual speaker recognition

Conclusion In this work, we introduced LASPA, a novel approach for language-agnostic speaker disentanglement using prefix-tuned cross-attention. Our method effectively separates speaker iden- tity from linguistic information, addressing key challenges in multi-lingual speaker recognition. By integrating prefix-tuning, we achieve efficient adaptation while...

work page
[1]

Introduction Speaker recognition in multi-lingual contexts presents sig- nificant challenges due to the complex interaction between language-specific characteristics, speaker identity, and acous- tic variability. When speakers use different languages, their phonetic and prosodic patterns are influenced by the linguis- tic structure of each language and ca...

work page
[2]

The paper proposes a joint learning-based approach to derive language-independent speaker embeddings for multi-lingual speaker recognition and speaker diarization

work page
[3]

Speaker and language encoders, along with prefix-tuners are employed to fuse embeddings, ensuring accurate signal re- construction by the decoder

work page
[4]

To enhance the learning of speaker embedding, this work uses multiple loss functions like Mean Absolute Pearson’s Correlation (MAPC) [9], Additive Angular Margin Softmax (AAM Softmax) [11], Mean Squared Error (MSE), and Nega- tive Log Likelihood (NLL)

work page
[5]

Previous Work Previous studies have shown that speaker representations are often entangled with factors such as emotions [12, 13], ac- cent [14], age [15], and environmental conditions like noise and reverberation [16]. More recently, research has validated that linguistic features embedded in speaker representations make them susceptible to language vari...

work page Pith review arXiv 2025
[6]

Model Architecture The architectural overview of the proposed system is shown in Fig. 1. The architecture comprises a Speaker Encoder, a Lan- guage Encoder, two Prefix-Tuners, and a Decoder. The input waveform is initially resampled to a16kHz sampling rate. Sub- sequently, mel-spectrogram is computed using a Hamming win- dow with a25ms window size and a10...

work page 2021
[7]

In the Speaker Encoder, the mel- spectrogram passes through a feature extraction module and a fully connected layer

Training and Inference During training, the input waveform is converted into mel- spectrograms, denoted asX i. In the Speaker Encoder, the mel- spectrogram passes through a feature extraction module and a fully connected layer. Details of the feature extractor can be found in Section 5. The output of the fully connected layer is the speaker embedding,E sp...

work page

Show all 40 references

[8]

Baselines We use ResNet-(S [17] and L [18]), ECAPA [19] and ReD- imNet [20] as speaker encoder architectures which are trained from scratch

Experimental Analysis 5.1. Baselines We use ResNet-(S [17] and L [18]), ECAPA [19] and ReD- imNet [20] as speaker encoder architectures which are trained from scratch. For the Language Encoder we use ECAPA archi- tecture and is initialized from the model pre-trained on V oxLin...

2021
[10]

Disentangled representation learning for multilingual speaker recognition,

K. Nam, Y . Kim, J. Huh, H.-S. Heo, J. weon Jung, and J. S. Chung, “Disentangled representation learning for multilingual speaker recognition,” inInterspeech 2023, 2023, pp. 5316–5320

2023
[11]

Generating multilingual voices using speaker space translation based on bilingual speaker data,

S. M. E. M. A. Conkie, “Generating multilingual voices using speaker space translation based on bilingual speaker data,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 7624–7628. [Online]. Available: https://arx...

2020 arXiv
[12]

Unsupervised domain adaptation by backpropagation,

Y . Ganin and V . Lempitsky, “Unsupervised domain adaptation by backpropagation,” inProceedings of the 32nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, F. Bach and D. Blei, Eds., vol. 37. Lille, France: PMLR, 07–09 Jul 2015, pp....

2015
[13]

Disentan- gled speaker representation learning via mutual information min- imization,

S. H. Mun, M. H. Han, M. Kim, D. Lee, and N. S. Kim, “Disentan- gled speaker representation learning via mutual information min- imization,” in2022 Asia-Pacific Signal and Information Process- ing Association Annual Summit and Conference (APSIPA ASC). IEEE, 2022, pp. 89–96

2022
[14]

Disentangled speaker embedding for ro- bust speaker verification,

L. Yi and M.-W. Mak, “Disentangled speaker embedding for ro- bust speaker verification,” inICASSP 2022 - 2022 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 7662–7666

2022
[15]

Deep Representation Decomposition for Rate-Invariant Speaker Verifi- cation,

F. Tong, S. Zheng, H. Zhou, X. Xie, Q. Hong, and L. Li, “Deep Representation Decomposition for Rate-Invariant Speaker Verifi- cation,” inProc. Speaker Odyssey, 2022, pp. 228–232

2022
[16]

Augmenta- tion adversarial training for self-supervised speaker representation learning,

J. Kang, J. Huh, H. S. Heo, and J. S. Chung, “Augmenta- tion adversarial training for self-supervised speaker representation learning,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1253–1262, 2022

2022
[17]

Towards principled methods for training generative adversarial networks,

M. Arjovsky and L. Bottou, “Towards principled methods for training generative adversarial networks,” inInternational Conference on Learning Representations, 2017. [Online]. Available: https://openreview.net/forum?id=Hk4 qw5xe

2017
[18]

Disentan- gled speaker and nuisance attribute embedding for robust speaker verification,

W. H. Kang, S. H. Mun, M. H. Han, and N. S. Kim, “Disentan- gled speaker and nuisance attribute embedding for robust speaker verification,”IEEE Access, vol. 8, pp. 141 838–141 849, 2020

2020
[19]

Disentangled represen- tation with dual-stage feature learning for face anti-spoofing,

Y .-C. Wang, C.-Y . Wang, and S.-H. Lai, “Disentangled represen- tation with dual-stage feature learning for face anti-spoofing,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 1955–1964. [Online]. Avail- able: https://openaccess...

2022
[20]

Arcface: Additive angular margin loss for deep face recognition,

J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” inProceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, 2019, pp. 4690–4699

2019
[21]

Disentangling style factors from speaker representations,

J. Williams and S. King, “Disentangling style factors from speaker representations,” inInterspeech 2019, 2019, pp. 3945–3949

2019
[22]

X- vectors meet emotions: A study on dependencies between emo- tion and speaker recognition,

R. Pappagari, T. Wang, J. Villalba, N. Chen, and N. Dehak, “X- vectors meet emotions: A study on dependencies between emo- tion and speaker recognition,” inICASSP 2020 - 2020 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 7169–7173

2020
[23]

Accent and speaker disentanglement in many- to-many voice conversion,

Z. Wang, W. Ge, X. Wang, S. Yang, W. Gan, H. Chen, H. Li, L. Xie, and X. Li, “Accent and speaker disentanglement in many- to-many voice conversion,” in2021 12th International Symposium on Chinese Spoken Language Processing (ISCSLP), 2021, pp. 1– 5

2021
[24]

Probing the information encoded in x-vectors,

D. Raj, D. Snyder, D. Povey, and S. Khudanpur, “Probing the information encoded in x-vectors,” in2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, Dec. 2019. [Online]. Available: http://dx.doi.org/10.1109/ ASRU46091.2019.9003979

2019
[25]

Speaker recognition: a tutorial,

J. Campbell, “Speaker recognition: a tutorial,”Proceedings of the IEEE, vol. 85, no. 9, pp. 1437–1462, 1997

1997
[26]

In defence of metric learning for speaker recognition,

J. S. Chung, J. Huh, S. Mun, M. Lee, H.-S. Heo, S. Choe, C. Ham, S. Jung, B.-J. Lee, and I. Han, “In defence of metric learning for speaker recognition,” inInterspeech 2020, 2020, pp. 2977–2981

2020
[27]

The ins and outs of speaker recognition: lessons from voxsrc 2020,

Y . Kwon, H.-S. Heo, B.-J. Lee, and J. S. Chung, “The ins and outs of speaker recognition: lessons from voxsrc 2020,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 5809–5813

2020
[28]

ECAPA- TDNN: Emphasized Channel Attention, Propagation and Ag- gregation in TDNN Based Speaker Verification,

B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA- TDNN: Emphasized Channel Attention, Propagation and Ag- gregation in TDNN Based Speaker Verification,” inInterspeech 2020, 2020, pp. 3830–3834

2020
[29]

Reshape dimensions network for speaker recognition,

I. Yakovlev, R. Makarov, A. Balykin, P. Malov, A. Okhotnikov, and N. Torgashov, “Reshape dimensions network for speaker recognition,” inInterspeech 2024, 2024, pp. 3235–3239

2024
[30]

Prefix-tuning: Optimizing continuous prompts for generation,

X. L. Li and P. Liang, “Prefix-tuning: Optimizing continuous prompts for generation,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). O...

2021
[31]

Cross-modal fusion and attention mechanism for weakly supervised video anomaly detection,

A. Ghadiya, P. Kar, V . Chudasama, and P. Wasnik, “Cross-modal fusion and attention mechanism for weakly supervised video anomaly detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 1965–1974. [Online]. Available: https: /...

2024
[32]

Atten- tion is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Atten- tion is all you need,” inAdvances in Neural Information Processing Systems, I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garn...

2017
[33]

Long Short-Term Memory,

S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 11 1997

1997
[34]

Multilayer perceptron and neural networks,

M.-C. Popescu, V . E. Balas, L. Perescu-Popescu, and N. Mas- torakis, “Multilayer perceptron and neural networks,”WSEAS Transactions on Circuits and Systems, vol. 8, no. 7, pp. 579–588, 2009

2009
[35]

V oxceleb2: Deep speaker recognition,

J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: Deep speaker recognition,” inInterspeech 2018, 2018, pp. 1086–1090

2018
[36]

V oxlingua107: A dataset for spoken lan- guage recognition,

J. Valk and T. Alum ¨ae, “V oxlingua107: A dataset for spoken lan- guage recognition,” in2021 IEEE Spoken Language Technology Workshop (SLT), 2021, pp. 652–658

2021
[37]

V oxceleb: A large- scale speaker identification dataset,

A. Nagrani, J. S. Chung, and A. Zisserman, “V oxceleb: A large- scale speaker identification dataset,” inInterspeech 2017, 2017, pp. 2616–2620

2017
[38]

Nisp: A multi-lingual multi-accent dataset for speaker profiling,

S. B. Kalluri, D. Vijayasenan, S. Ganapathy, R. R. M, and P. Kr- ishnan, “Nisp: A multi-lingual multi-accent dataset for speaker profiling,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 6953–6957

2021
[39]

The displace challenge 2023 - diarization of speaker and language in conversational environments,

S. Baghel, S. Ramoji, Sidharth, R. H, P. Singh, S. Jain, P. Roy Chowdhuri, K. Kulkarni, S. Padhi, D. Vijayasenan, and S. Gana- pathy, “The displace challenge 2023 - diarization of speaker and language in conversational environments,” inInterspeech 2023, 2023, pp. 3562–3566

2023
[40]

Ensemble of incremental system enhancements for robust speaker diariza- tion in code-switched real-life audios,

R. Gohil, R. Viswanathan, S. Agrawal, C. M. Vikram, M. R. Kam- ble, K. Sabu, M. A. B. Shaik, and K. K. S. Rajesh, “Ensemble of incremental system enhancements for robust speaker diariza- tion in code-switched real-life audios,” inSpeech and Computer, A. Karpov, K. Samudravijay...

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [9]

Our method effectively separates speaker iden- tity from linguistic information, addressing key challenges in multi-lingual speaker recognition

Conclusion In this work, we introduced LASPA, a novel approach for language-agnostic speaker disentanglement using prefix-tuned cross-attention. Our method effectively separates speaker iden- tity from linguistic information, addressing key challenges in multi-lingual speaker recognition. By integrating prefix-tuning, we achieve efficient adaptation while...

work page

[2] [1]

Introduction Speaker recognition in multi-lingual contexts presents sig- nificant challenges due to the complex interaction between language-specific characteristics, speaker identity, and acous- tic variability. When speakers use different languages, their phonetic and prosodic patterns are influenced by the linguis- tic structure of each language and ca...

work page

[3] [2]

The paper proposes a joint learning-based approach to derive language-independent speaker embeddings for multi-lingual speaker recognition and speaker diarization

work page

[4] [3]

Speaker and language encoders, along with prefix-tuners are employed to fuse embeddings, ensuring accurate signal re- construction by the decoder

work page

[5] [4]

To enhance the learning of speaker embedding, this work uses multiple loss functions like Mean Absolute Pearson’s Correlation (MAPC) [9], Additive Angular Margin Softmax (AAM Softmax) [11], Mean Squared Error (MSE), and Nega- tive Log Likelihood (NLL)

work page

[6] [5]

Previous Work Previous studies have shown that speaker representations are often entangled with factors such as emotions [12, 13], ac- cent [14], age [15], and environmental conditions like noise and reverberation [16]. More recently, research has validated that linguistic features embedded in speaker representations make them susceptible to language vari...

work page Pith review arXiv 2025

[7] [6]

Model Architecture The architectural overview of the proposed system is shown in Fig. 1. The architecture comprises a Speaker Encoder, a Lan- guage Encoder, two Prefix-Tuners, and a Decoder. The input waveform is initially resampled to a16kHz sampling rate. Sub- sequently, mel-spectrogram is computed using a Hamming win- dow with a25ms window size and a10...

work page 2021

[8] [7]

In the Speaker Encoder, the mel- spectrogram passes through a feature extraction module and a fully connected layer

Training and Inference During training, the input waveform is converted into mel- spectrograms, denoted asX i. In the Speaker Encoder, the mel- spectrogram passes through a feature extraction module and a fully connected layer. Details of the feature extractor can be found in Section 5. The output of the fully connected layer is the speaker embedding,E sp...

work page