pith. sign in

arxiv: 2606.10654 · v1 · pith:VM6FLU55new · submitted 2026-06-09 · 💻 cs.CL

Speaker Group Encoding in Self-supervised Speech Recognition Models

Felix Herron , Solange Rossato Alexandre Allauzen , Benoit Favre , Fran\c{c}ois Portet This is my paper

Pith reviewed 2026-06-27 13:13 UTC · model grok-4.3

classification 💻 cs.CL
keywords self-supervised speech modelsspeaker group informationphonetic versus semantic variancespeaker identification finetuningASR finetuningfairness in speech recognitionembedding analysis
0
0 comments X

The pith

Self-supervised speech models encode speaker groups such as gender, age, dialect, ethnicity and native status, with finetuning for speaker identification amplifying phonetic aspects while ASR finetuning discards them.

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

The paper tests what self-supervised speech recognition models learn about categories of speakers. It compares pretrained models to versions finetuned on speaker identification, on automatic speech recognition, and on ASR with fairness adjustments. The models turn out to carry information about gender, age, dialect, ethnicity, and whether a speaker is native. Finetuning changes which parts of this information stay or grow stronger, depending on whether the group differences arise mostly from sound patterns or from meaning patterns. The work tracks how this information appears across layers and inside specific parts of the embeddings.

Core claim

Self-supervised speech models encode information about speaker group categories including gender, age, dialect, ethnicity, and native speaker status. Finetuning for speaker identification amplifies speaker group information whose variance is more phonetic in nature but does not amplify information whose variance is more semantic in nature. Finetuning for automatic speech recognition discards phonetically variant speaker group information but retains semantically variant speaker group information. Fairness-enhancing ASR algorithms change the amount of speaker group information encoded, mainly for phonetically variant categories.

What carries the argument

Speaker group information (SGI) tracked through model embeddings, divided into phonetically variant versus semantically variant categories and measured before and after different finetuning tasks.

If this is right

  • Finetuning on speaker identification increases the strength of phonetically driven group signals inside the model.
  • Finetuning on automatic speech recognition removes phonetically driven group signals while leaving semantically driven ones in place.
  • Fairness adjustments to ASR training mainly reduce phonetic group signals and leave semantic group signals largely unchanged.
  • Different speaker group categories are carried by distinct subdimensions of the embeddings and by different layers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Fairness methods may need separate handling for semantic speaker information to achieve uniform reduction across all group categories.
  • Task-specific control of which speaker signals to retain could be added during pretraining rather than only at finetuning time.
  • The phonetic-versus-semantic split may appear in other self-supervised audio models when similar finetuning experiments are run.

Load-bearing premise

Speaker group categories can be correctly split into those whose main differences are phonetic versus those whose main differences are semantic, and this split explains why finetuning affects them differently.

What would settle it

Re-label the speaker groups by direct acoustic measurements of phonetic variance versus semantic variance and check whether the amplification and discarding patterns still align with the speaker-ID versus ASR finetuning results.

Figures

Figures reproduced from arXiv: 2606.10654 by Benoit Favre, Felix Herron, Fran\c{c}ois Portet, Solange Rossato Alexandre Allauzen.

Figure 1
Figure 1. Figure 1: SGI captured by S3Ms, measured using linear probes, as described in Equation 1. vs any individual SGI probing task. Despite having over 1000 classes (vs a handful for the SGI tasks), our probes achieved almost perfect performance for SID, while they were nowhere near perfect for any SGC besides gender. 4.1 Impact of model pretraining on SG encoding The pretrained model architectures (black) which captures … view at source ↗
Figure 2
Figure 2. Figure 2: Cosine similarity between principal components of each SGC compared with principal [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cosine similarity between principal components of each SGC compared with principal [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Earlier frames of a pretrained S3M cap￾ture more SGI than later frames, though the entire frame captures most of all. 5.2 Different SGCs are encoded across different latent dimensions [15] show that SID information and phonetic information are stored along orthogonal dimensions within latent embeddings of S3Ms. They show this by comparing the principal components of the centroid matrix for each class, in t… view at source ↗
read the original abstract

We investigate what self-supervised speech recognition models (S3Ms) learn about speaker groups (SGs). We examine several states of S3Ms: pretrained, finetuned on speaker identification (SID), finetuned on automatic speech recognition (ASR), and ASR-finetuned using a fairness enhancing algorithm. We find that S3Ms encode information about several speaker group categories (SGCs), including their gender, age, dialect, ethnicity, and whether they are a native speaker. We find that finetuning for SID amplifies certain SGCs, namely those whose variance is more phonetic in nature, though it does not amplify other SGCs, namely those whose variance is more semantic in nature. On the other hand, finetuning for ASR discards phonetically variant speaker group information (SGI) but retains semantically variant SGI. We find that ASR algorithms designed for fairness improvement change to what extent SGI is encoded in S3Ms; however, this is primarily true for for phonetically variant SGCs, and less true for semantically variant SGCs. We discuss how SGI is encoded by each layer, and identify subdimensions of embeddings responsible for encoding different SGCs. Finally, we discuss how our findings could be beneficial in designing fairer ASR algorithms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper investigates speaker group encoding in self-supervised speech recognition models (S3Ms). It compares pretrained models to those finetuned on speaker identification (SID), automatic speech recognition (ASR), and fairness-enhanced ASR. The central claims are that S3Ms encode information about speaker group categories (SGCs) including gender, age, dialect, ethnicity, and native-speaker status; that SID finetuning amplifies SGCs whose variance is primarily phonetic while leaving semantically variant SGCs unchanged; that ASR finetuning discards phonetically variant speaker group information (SGI) but retains semantically variant SGI; and that fairness algorithms primarily affect phonetic SGCs. Additional analyses cover layer-wise encoding and subdimensions of embeddings responsible for different SGCs.

Significance. If the phonetic-versus-semantic distinction can be independently validated and the probing results replicated with appropriate controls, the work would provide actionable measurements for controlling unwanted speaker attributes during ASR finetuning and for designing fairness interventions. The multi-state comparison (pretrained, SID, ASR, fairness) supplies concrete empirical data that could inform model development.

major comments (2)
  1. [Results and Discussion sections describing SGC categorization and finetuning effects] The assignment of SGCs to 'phonetic variance' (gender, age, dialect) versus 'semantic variance' (ethnicity, native-speaker status) is load-bearing for the interpretation of all finetuning effects, yet no explicit criteria, quantitative proxy (e.g., mutual information with formant or lexical features), pre-specified validation, or citation to independent classification is supplied. Without such grounding the differential amplification/discarding claims cannot be evaluated as causal rather than post-hoc.
  2. [Methods and Experimental Setup] The abstract and methods description supply no information on the datasets, number of speakers per SGC, probing classifier architecture, statistical tests, or controls for confounding acoustic or lexical factors. These omissions prevent assessment of whether the reported encoding differences are robust.
minor comments (2)
  1. [Methods] Clarify the operationalization of 'phonetic' versus 'semantic' variance for each SGC with a table or explicit decision rule.
  2. [Layer-wise analysis] Add layer indices and embedding dimensions when discussing subdimensions responsible for specific SGCs.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address the two major comments point by point below. Where the comments identify gaps in justification or reporting, we agree that revisions are warranted and will incorporate the requested clarifications.

read point-by-point responses

  1. Referee: [Results and Discussion sections describing SGC categorization and finetuning effects] The assignment of SGCs to 'phonetic variance' (gender, age, dialect) versus 'semantic variance' (ethnicity, native-speaker status) is load-bearing for the interpretation of all finetuning effects, yet no explicit criteria, quantitative proxy (e.g., mutual information with formant or lexical features), pre-specified validation, or citation to independent classification is supplied. Without such grounding the differential amplification/discarding claims cannot be evaluated as causal rather than post-hoc.

    Authors: We acknowledge that the phonetic-versus-semantic distinction requires more explicit grounding than is currently provided. The categorization was motivated by established linguistic distinctions: attributes such as gender, age, and dialect primarily modulate acoustic-phonetic realization (e.g., formant shifts, prosody), whereas ethnicity and native-speaker status more strongly influence lexical choice and semantic content in the datasets examined. However, we agree that this rationale should be stated with criteria and supporting citations rather than left implicit. In the revised manuscript we will add a dedicated subsection (likely in Methods or an expanded Results) that (i) states the pre-specified criteria used, (ii) supplies relevant linguistic references, and (iii) reports a simple quantitative check (e.g., correlation of each SGC with selected acoustic features). This will allow readers to evaluate the distinction independently. revision: yes

  2. Referee: [Methods and Experimental Setup] The abstract and methods description supply no information on the datasets, number of speakers per SGC, probing classifier architecture, statistical tests, or controls for confounding acoustic or lexical factors. These omissions prevent assessment of whether the reported encoding differences are robust.

    Authors: We agree that the current Methods section is insufficiently detailed for reproducibility and evaluation. Although the full manuscript contains the underlying data sources and probing setup, these elements are not presented with the required specificity. In the revision we will expand the Methods section to include: (a) exact dataset names and splits with speaker counts per SGC, (b) the architecture and hyperparameters of the probing classifiers, (c) the statistical tests and multiple-comparison corrections applied, and (d) any acoustic or lexical controls (or explicit statement that none were used). These additions will directly address the referee’s concern about robustness. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical probing with no derivations or self-referential fits

full rationale

The work reports measurements of speaker group information in pretrained and finetuned S3Ms via probing classifiers. No equations, parameter fittings, or predictions are defined that reduce to the input data by construction. The phonetic-vs-semantic variance distinction is an interpretive label applied to observed categories; it does not appear as a self-definitional step, a fitted input renamed as prediction, or a load-bearing self-citation. The analysis is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper performs empirical analysis of pretrained and finetuned models and introduces no new free parameters, mathematical axioms, or postulated entities.

pith-pipeline@v0.9.1-grok · 5773 in / 1150 out tokens · 24125 ms · 2026-06-27T13:13:50.770780+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 10 canonical work pages

  1. [1]

    Similarities between Arabic dialects: Investigating geographical proximity.Information Processing & Management, 59(1):102770, January

    Abdulkareem Alsudais, Wafa Alotaibi, and Faye Alomary. Similarities between Arabic dialects: Investigating geographical proximity.Information Processing & Management, 59(1):102770, January

  2. [2]

    doi: 10.1016/j.ipm.2021.102770

    ISSN 0306-4573. doi: 10.1016/j.ipm.2021.102770

    work page doi:10.1016/j.ipm.2021.102770 2021

  3. [3]

    Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations, October 2020

    Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations, October 2020

    2020

  4. [4]

    WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing , volume=

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Xiangzhan Yu, and Furu Wei. WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing.IEEE Journal of Selected Topics in Signal Pr...

    work page doi:10.1109/jstsp.2022.3188113 2022

  5. [5]

    Self-Supervised Speech Representations are More Phonetic than Semantic, June 2024

    Kwanghee Choi, Ankita Pasad, Tomohiko Nakamura, Satoru Fukayama, Karen Livescu, and Shinji Watanabe. Self-Supervised Speech Representations are More Phonetic than Semantic, June 2024

    2024

  6. [6]

    Similarity Analysis of Self-Supervised Speech Representations, February 2021

    Yu-An Chung, Yonatan Belinkov, and James Glass. Similarity Analysis of Self-Supervised Speech Representations, February 2021

    2021

  7. [7]

    Proximité rythmique entre apprenants et natifs du français Évaluation d’une métrique basée sur le CEFC

    Sylvain Coulange and Solange Rossato. Proximité rythmique entre apprenants et natifs du français Évaluation d’une métrique basée sur le CEFC. In Christophe Benzitoun, Chloé Braud, Laurine Huber, David Langlois, Slim Ouni, Sylvain Pogodalla, and Stéphane Schneider, editors, Actes de La 6e Conférence Conjointe Journées d’Études Sur La Parole (JEP, 33e Éditi...

    2020

  8. [8]

    Dialect-Specific Models for Automatic Speech Recognition of African American Vernacular English

    Rachel Dorn. Dialect-Specific Models for Automatic Speech Recognition of African American Vernacular English. In Venelin Kovatchev, Irina Temnikova, Branislava Šandrih, and Ivelina Nikolova, editors,Proceedings of the Student Research Workshop Associated with RANLP 2019, pages 16–20, Varna, Bulgaria, September 2019. INCOMA Ltd. doi: 10.26615/issn.2603-282...

    work page doi:10.26615/issn.2603-2821.2019_003 2019

  9. [9]

    LeBenchmark: A Reproducible Framework for Assessing Self-Supervised Representation Learning from Speech

    Solene Evain, Ha Nguyen, Hang Le, Marcely Zanon Boito, Salima Mdhaffar, Sina Alisamir, Ziyi Tong, Natalia Tomashenko, Marco Dinarelli, Titouan Parcollet, Alexandre Allauzen, Yannick Esteve, Benjamin Lecouteux, Francois Portet, Solange Rossato, Fabien Ringeval, Didier Schwab, and Laurent Besacier. LeBenchmark: A Reproducible Framework for Assessing Self-Su...

    work page doi:10.21437/interspeech.2021-556 2021

  10. [10]

    Silence is Sweeter Than Speech: Self- Supervised Model Using Silence to Store Speaker Information, May 2022

    Chi-Luen Feng, Po-chun Hsu, and Hung-yi Lee. Silence is Sweeter Than Speech: Self- Supervised Model Using Silence to Store Speaker Information, May 2022

    2022

  11. [11]

    Towards inclusive automatic speech recognition.Computer Speech & Language, 84:101567, March 2024

    Siyuan Feng, Bence Mark Halpern, Olya Kudina, and Odette Scharenborg. Towards inclusive automatic speech recognition.Computer Speech & Language, 84:101567, March 2024. ISSN 0885-2308. doi: 10.1016/j.csl.2023.101567

    work page doi:10.1016/j.csl.2023.101567 2024

  12. [12]

    Domain-Adversarial Training of Neural Networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-Adversarial Training of Neural Networks. In Gabriela Csurka, editor,Domain Adaptation in Computer Vision Applications, pages 189–209. Springer International Publishing, Cham, 2017. ISBN 978-3-319-58346-4 978-3-...

    work page doi:10.1007/978-3-319-58347-1_10 2017

  13. [13]

    HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units, June 2021

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units, June 2021

    2021

  14. [14]

    Accent-Robust Automatic Speech Recognition Using Supervised and Unsupervised Wav2vec Embeddings, October 2021

    Jialu Li, Vimal Manohar, Pooja Chitkara, Andros Tjandra, Michael Picheny, Frank Zhang, Xiaohui Zhang, and Yatharth Saraf. Accent-Robust Automatic Speech Recognition Using Supervised and Unsupervised Wav2vec Embeddings, October 2021

    2021

  15. [15]

    Nuanced metrics for measuring unintended bias with real data for text classification

    Lanna Lima, Vasco Furtado, Elizabeth Furtado, and Virgilio Almeida. Empirical Analysis of Bias in V oice-based Personal Assistants. InCompanion Proceedings of The 2019 World Wide Web Conference, pages 533–538, San Francisco USA, May 2019. ACM. ISBN 978-1-4503-6675-5. doi: 10.1145/3308560.3317597

    work page doi:10.1145/3308560.3317597 2019

  16. [16]

    Self-supervised Predictive Coding Models Encode Speaker and Phonetic Information in Orthogonal Subspaces, December 2023

    Oli Liu, Hao Tang, and Sharon Goldwater. Self-supervised Predictive Coding Models Encode Speaker and Phonetic Information in Orthogonal Subspaces, December 2023

    2023

  17. [17]

    Orthogonality and isotropy of speaker and phonetic information in self-supervised speech representations, June 2024

    Mukhtar Mohamed, Oli Danyi Liu, Hao Tang, and Sharon Goldwater. Orthogonality and isotropy of speaker and phonetic information in self-supervised speech representations, June 2024

    2024

  18. [18]

    Layer-wise Analysis of a Self-supervised Speech Representation Model, December 2022

    Ankita Pasad, Ju-Chieh Chou, and Karen Livescu. Layer-wise Analysis of a Self-supervised Speech Representation Model, December 2022

    2022

  19. [19]

    Comparative layer-wise analysis of self- supervised speech models, March 2023

    Ankita Pasad, Bowen Shi, and Karen Livescu. Comparative layer-wise analysis of self- supervised speech models, March 2023

    2023

  20. [20]

    What Do Self-Supervised Speech Models Know About Words?, January 2024

    Ankita Pasad, Chung-Ming Chien, Shane Settle, and Karen Livescu. What Do Self-Supervised Speech Models Know About Words?, January 2024

    2024

  21. [21]

    SpeechBrain: A General-Purpose Speech Toolkit, June 2021

    Mirco Ravanelli, Titouan Parcollet, Peter Plantinga, Aku Rouhe, Samuele Cornell, Loren Lugosch, Cem Subakan, Nauman Dawalatabad, Abdelwahab Heba, Jianyuan Zhong, Ju-Chieh Chou, Sung-Lin Yeh, Szu-Wei Fu, Chien-Feng Liao, Elena Rastorgueva, François Grondin, William Aris, Hwidong Na, Yan Gao, Renato De Mori, and Yoshua Bengio. SpeechBrain: A General-Purpose...

    2021

  22. [22]

    Linear Adversarial Concept Erasure, December 2024

    Shauli Ravfogel, Michael Twiton, Yoav Goldberg, and Ryan Cotterell. Linear Adversarial Concept Erasure, December 2024

    2024

  23. [23]

    Analyzing Acoustic Word Embeddings from Pre-trained Self-supervised Speech Models, March 2023

    Ramon Sanabria, Hao Tang, and Sharon Goldwater. Analyzing Acoustic Word Embeddings from Pre-trained Self-supervised Speech Models, March 2023. 9

    2023

  24. [24]

    Sonos V oice Control Bias Assessment Dataset: A Methodology for Demographic Bias Assessment in V oice Assistants

    Chloe Sekkat, Fanny Leroy, Salima Mdhaffar, Blake Perry Smith, Yannick Estève, Joseph Dureau, and Alice Coucke. Sonos V oice Control Bias Assessment Dataset: A Methodology for Demographic Bias Assessment in V oice Assistants. In Nicoletta Calzolari, Min-Yen Kan, Veronique Hoste, Alessandro Lenci, Sakriani Sakti, and Nianwen Xue, editors,Proceedings of the...

    2024

  25. [25]

    2020, in ICASSP 2020, 1434–1438, doi: 10.1109/ICASSP40776.2020.9053284

    Suwon Shon, Ahmed Ali, Younes Samih, Hamdy Mubarak, and James Glass. ADI17: A Fine-Grained Arabic Dialect Identification Dataset. InICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8244–8248, Barcelona, Spain, May 2020. IEEE. ISBN 978-1-5090-6631-5. doi: 10.1109/ICASSP40776.2020.9052982

    work page doi:10.1109/icassp40776.2020.9052982 2020

  26. [26]

    X- Vectors: Robust DNN Embeddings for Speaker Recognition

    David Snyder, Daniel Garcia-Romero, Gregory Sell, Daniel Povey, and Sanjeev Khudanpur. X- Vectors: Robust DNN Embeddings for Speaker Recognition. In2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5329–5333, April 2018. doi: 10.1109/ ICASSP.2018.8461375

    arXiv 2018

  27. [27]

    Analyzing Speaker Information in Self-Supervised Models to Improve Zero-Resource Speech Processing, August 2021

    Benjamin van Niekerk, Leanne Nortje, Matthew Baas, and Herman Kamper. Analyzing Speaker Information in Self-Supervised Models to Improve Zero-Resource Speech Processing, August 2021

    2021

  28. [28]

    Irina-Elena Veliche, Zhuangqun Huang, Vineeth Ayyat Kochaniyan, Fuchun Peng, Ozlem Kalinli, and Michael L. Seltzer. Towards measuring fairness in speech recognition: Fair-Speech dataset, August 2024

    2024

  29. [29]

    Open Implementation and Study of BEST-RQ for Speech Processing, September 2024

    Ryan Whetten, Titouan Parcollet, Marco Dinarelli, and Yannick Estève. Open Implementation and Study of BEST-RQ for Speech Processing, September 2024

    2024

  30. [30]

    Doyle, Leigh Clark, and Benjamin R

    Yunhan Wu, Daniel Rough, Anna Bleakley, Justin Edwards, Orla Cooney, Philip R. Doyle, Leigh Clark, and Benjamin R. Cowan. See What I’m Saying? Comparing Intelligent Personal Assistant Use for Native and Non-Native Language Speakers. In22nd International Conference on Human-Computer Interaction with Mobile Devices and Services, MobileHCI ’20, pages 1–9, Ne...

    work page doi:10.1145/3379503.3403563 2020

  31. [31]

    Jeff Lai, Kushal Lakhotia, Yist Y

    Shu-wen Yang, Po-Han Chi, Yung-Sung Chuang, Cheng-I. Jeff Lai, Kushal Lakhotia, Yist Y . Lin, Andy T. Liu, Jiatong Shi, Xuankai Chang, Guan-Ting Lin, Tzu-Hsien Huang, Wei-Cheng Tseng, Ko-tik Lee, Da-Rong Liu, Zili Huang, Shuyan Dong, Shang-Wen Li, Shinji Watanabe, Abdelrahman Mohamed, and Hung-yi Lee. SUPERB: Speech processing Universal PERformance Benchm...

    2021

  32. [32]

    Less Forgetting for Better Generalization: Exploring Continual-learning Fine-tuning Methods for Speech Self-supervised Representations, June 2024

    Salah Zaiem, Titouan Parcollet, and Slim Essid. Less Forgetting for Better Generalization: Exploring Continual-learning Fine-tuning Methods for Speech Self-supervised Representations, June 2024

    2024

  33. [33]

    In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Wei Zhou, Haotian Wu, Jingjing Xu, Mohammad Zeineldeen, Christoph Lüscher, Ralf Schlüter, and Hermann Ney. Enhancing and Adversarial: Improve ASR with Speaker Labels. InICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, June 2023. doi: 10.1109/ICASSP49357.2023.10096722. 10

    work page doi:10.1109/icassp49357.2023.10096722 2023