Pith. sign in

REVIEW 3 major objections 5 minor 61 references

Different Speech Translation Models Encode and Translate Speaker Gender Differently

T0 review · 3 major / 5 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 This paper shows that speech translation architecture determines whether a model encodes speaker gender, with adapter-based speech+MT models erasing most of it, and that gender translation accuracy tracks encoded gender strength…

desk verdict Real architecture-level difference in gender encoding, but 'do not encode' overstates probe failure without sensitivity controls. read the letter →

arxiv 2506.02172 v1 pith:KP3LWCLE submitted 2025-06-02 cs.CL

Dennis Fucci , Marco Gaido , Matteo Negri , Luisa Bentivogli , Andre Martins , Giuseppe Attanasio This is my paper
classification cs.CL
keywords speechtranslationgenderencodingprobingadaptersbiasmasculinedefaultinterpretabilityspeaker
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

Across English-to-French/Italian/Spanish speech translation, the paper finds that architecture decides whether a model can hear the speaker's gender. A traditional encoder-decoder trained end-to-end on speech translation keeps gender in its hidden states, so a probe can read it with high accuracy and translations usually inflect correctly. Newer speech+MT systems, which use a pretrained speech encoder feeding a machine translation model through an adapter, shed most of that gender information at the adapter, and their translations default to masculine forms when the speaker is female. The paper reports a near-perfect correlation ($R^2=0.99$) between how strongly a model's representations encode gender and how accurately it translates gender-marked words for the speaker. A sympathetic reader would take this as evidence that acoustic gender cues are usable by speech translation systems, that adapters are the weak point, and that scrubbing gender information can worsen, not improve, fair translation.

What carries the argument

The load-bearing instrument is an attention-based probe: a single learnable query attends over the model's hidden-state sequence (Q-Former style) and a linear layer classifies the pooled representation into She/He. It outperforms mean and max pooling and positional sampling, and its attention weights reveal where gender lives in the sequence (mostly early positions). The second piece of machinery is the architectural comparison itself: probing the same speech encoder before and after the adapter isolates where gender information is lost in speech+MT models, and correlating probing F1 with MuST-SHE gender accuracy turns encoding strength into a predictor of translation behaviour.

What would settle it

Run Hewitt–Liang control probes on the post-adapter representations: if the control probes score as low as the real probes, the erased-gender conclusion is undermined, because the probe itself may be blind to the adapter's geometry rather than the gender being absent.

Watch

Extended reading notes

Core claim

The paper's central claim is that gender encoding is not a universal property of speech translation models but is shaped by architecture. On the same test data, probes reach roughly 92–96 macro F1 on the encoder outputs of a Transformer encoder-decoder (Wang et al., 2020), yet only 55–62 F1 on the post-adapter representations of SeamlessM4T and ZeroSwot; probing before the adapters shows substantially higher scores, so the adapter itself is implicated in removing gender information. The same ranking holds for gender translation: the encoder-decoder reaches about 86% average accuracy on MuST-SHE first-person references, while the adapter-based systems score about 53–63%, with feminine forms suffering most. The paper ties the two together with a linear fit ($R^2=0.99$, $p<0.01$) on the test-speaker set, arguing that when gender is weakly encoded the model falls back on a masculine default.

Load-bearing premise

The load-bearing premise is that a probe's low accuracy on a representation proves that the representation does not encode gender, rather than that the probe cannot extract it from that geometry.

Editorial extensions

If this is right

  • If the correlation holds, probing F1 can serve as a cheap diagnostic for a speech translation model's likely gender behaviour before running full translation evaluations.
  • Adapter design and training, not just the speech encoder, determine whether paralinguistic information such as gender survives into the translation module.
  • Removing gender information from representations is not a neutral debiasing step here; it is associated with a stronger masculine default and worse feminine-form accuracy.
  • Speech translation models can use acoustic cues for gender assignment when no other signal is available, but linguistic biases can still override those cues even when gender is well encoded.
  • Gender information is concentrated in early sequence positions, matching findings in ASR models, which may point to where architectural changes would have the most effect.

Reading between the lines

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

  • A direct test of the adapter-erasure story would be control probes (e.g., Hewitt–Liang) on post-adapter representations: if control probes fail as badly as real probes, the drop may reflect probe incompatibility with the adapter's embedding geometry rather than absent information.
  • The $R^2=0.99$ result covers three language pairs and three models; extending to more adapters, languages with different grammatical-gender systems, and LLM-based speech translation would show whether the correlation is a general law or a property of these particular checkpoints.
  • If preserving gender information improves accuracy, system designers face a trade-off between accuracy for female speakers and privacy and self-identification concerns, since acoustic gender inference can misgender transgender, non-binary, and vocally atypical speakers.
  • Amnesic probing, removing gender from encoder-decoder representations and watching translation accuracy drop, would strengthen the causal reading the paper draws from correlation.
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.

Referee Report

3 major / 5 minor

Summary. The paper probes gender encoding in the hidden states of three speech translation systems (a Transformer encoder-decoder, SeamlessM4T, and ZeroSwot) using a newly proposed attention-based probe, and links probing performance to gender translation accuracy on MuST-SHE first-person test items for English-to-French/Italian/Spanish. It reports that the traditional encoder-decoder encodes speaker gender with macro F1 up to 96, while the speech+MT models show much lower post-adapter F1 scores (roughly 55-61), that adapters remove a large share of gender information (pre-adapter vs post-adapter drops), and that probe F1 correlates strongly with gender translation accuracy (R²=0.99), with low-encoding systems exhibiting a masculine default.

Significance. If the findings hold, they constitute a valuable contribution to speech translation interpretability: they connect an internal-representation analysis to a concrete downstream bias, identify architecture-dependent variation in gender encoding, and provide a reusable attention-based probe. The paper also ships code under a permissive license, includes a careful manual analysis of out-of-coverage translations (Appendix C), and engages thoughtfully with the ethical implications in Section 8. However, the main claims currently rest on probe sensitivity assumptions and on summary statistics that lack uncertainty quantification, so the conclusions are not yet fully supported.

major comments (3)
  1. [Section 4.1, Table 1] The interpretation that Seamless and ZeroSwot post-adapter states "do not encode" gender (abstract) or encode it "minimally" is not supported without a probe sensitivity control. The reported post-adapter macro F1 values (55.6-60.7 on the balanced test-speaker set, with N≈550 per language) are above the 50% chance level, so the data are more consistent with weak but detectable encoding than with absence of encoding. Low F1 can also arise because the attention probe is less able to extract information from the post-adapter geometry (which has different dimensionality and distribution after compression) rather than because the information is absent. I recommend adding Hewitt-Liang control probes or a synthetic-signal positive control (for example, injecting a known gender-correlated signal into post-adapter states) to calibrate the probe, and reporting repeated-seed means and confidence intervals for the pre/post-adapter F1 drops before claiming that adapters "significantly remove" gender information.
  2. [Section 4.2, Figure 1] The central correlation claim (R²=0.99, p<0.01) is based on a small number of model-language points (3 models × 3 languages = 9) with no confidence intervals, no leave-one-out analysis, and no control for confounds such as overall model quality or language direction; a single influential point could drive the near-perfect fit. Since the probe scores used as the independent variable inherit the sensitivity problem described above, the correlation cannot by itself establish that "higher gender encoding capability leads to more accurate gender translation." The paper acknowledges in Limitations that correlation does not imply causation, but the abstract and conclusion state the causal implication more strongly; I recommend rephrasing these claims and adding a small-sample robustness check or per-language correlation analysis.
  3. [Section 3 (Data) / Table 5] The probing setup appears to train on the MuST-C en→es section only, yet test-speaker probing results are reported for en→fr and en→it in Table 1. If the same probe is evaluated on representations from language directions it never saw during training, the paper must validate cross-lingual transfer of the probe; hidden-state distributions can differ across target languages in multilingual models. If separate probes were trained per language, this should be stated explicitly and the per-language training sets and dimensions should be reported. This point matters because the language-average F1 scores and the Figure 1 correlation aggregate across these three conditions.
minor comments (5)
  1. [Table 2] In the Seamless average row, the average He coverage is listed as 42.61, but the three He coverage values (70.56, 59.66, 57.52) average to 62.58; this appears to be a typo.
  2. [Appendix E] Appendix E refers to "Figure 1" for the attention-weight distribution, but Figure 1 in Section 4.2 is the correlation plot; the appendix figure should be numbered separately (e.g., Figure 2).
  3. [Section 3] The text says test-speaker "contains over 550 samples and 1.30 hours of audio," but Table 5 lists three language-specific test-speaker sets totaling about 1700 samples and about 4 hours; please clarify whether the sentence refers to each language direction individually.
  4. [Section 4.1] The sentence "not exceeding 59.76 on test-generic and 61.80 on test-speaker (en→es)" is inaccurate for ZeroSwot post-adapter, whose en→fr test-speaker F1 is 61.80; the sentence should refer to the maximum overall or be rephrased.
  5. [Section 2] The key and value matrices are defined with K,V ∈ R^{L×d}, but the dimensions of the learnable query and the linear classification layer are not stated; adding these dimensions would clarify the architecture.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: probing and translation metrics are independently measured, the R²=0.99 correlation is descriptive rather than by-construction, and self-citations are contextual rather than load-bearing.

full rationale

I find no derivation step that reduces to its own inputs. The central empirical chain is: train an attention-based probe on hidden states to predict speaker gender (Section 2), measure probe F1 across model states (Section 4.1), measure gender translation accuracy with MuST-SHE (Section 4.2), and then correlate the two independently measured quantities; the reported R²=0.99 describes a linear association over 9 observed points, not a quantity forced by construction. No parameter is fitted to gender translation accuracy and then renamed a prediction: the probe is trained only on gender labels, and the regression in Figure 1 is descriptive. Self-citations to Attanasio et al. (2024), Bentivogli et al. (2020), and Gaido et al. (2020) supply datasets, prior observations, and annotation conventions, but the paper's architectural comparison and the probe-translation correlation rely on the paper's own measurements, not on those citations. The skeptical concern that post-adapter F1 values around 55-61 are interpreted as 'do not encode' without Hewitt-Liang control probes is a validity or sensitivity limitation, not circularity: the probe is a separate classifier, and low F1 is an observed result rather than an assumption embedded in the definition of the conclusion. The paper itself flags the probing paradigm's limits in Section 7 ('it has been criticized for not directly confirming whether the model uses the extracted information') and explicitly notes that 'correlation does not imply causation,' so the residual concern is a matter of evidential strength, not circular derivation.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central claims rest on domain assumptions about binary gender labels, probe validity, the absence of non-acoustic cues in the test set, and the representativeness of three models. No invented physical or theoretical entities are introduced; the attention-based probe is a methodology, not a postulated object.

free parameters (1)
  • Linear fit of probing F1 vs. gender accuracy (Figure 1) = R2=0.99; slope and intercept not reported
    This descriptive regression is used to support the claim that higher encoding causes more accurate translation. It is fit to a small set of model-language points and is not validated out-of-sample, so it functions as a fitted summary rather than a predictive derivation.
assumptions (5)
  • domain assumption Gender is treated as binary and is labeled from self-declared pronouns (She/He), with the audio signal assumed to contain voice features correlated with these labels.
    Used to define probe labels and the MuST-SHE evaluation. The authors acknowledge in Section 8 that gender is a continuum and that binary labeling is a simplification due to data availability.
  • domain assumption A high-accuracy probe indicates that the representation encodes gender, and a low-accuracy probe indicates that it does not.
    This is the core interpretive assumption in Section 4.1. It is known in the probing literature that probe failure can reflect probe insensitivity rather than absence of information, a concern the paper acknowledges but does not address with a control task.
  • domain assumption In test-speaker (MuST-SHE first-person items), the speaker's voice is the only cue for determining gender translation.
    Needed to interpret gender accuracy differences as evidence that models rely on acoustic cues. If additional non-acoustic contextual cues are present, the link between encoding and translation accuracy is weakened.
  • domain assumption The three selected models are representative of their architectural classes (encoder-decoder vs. speech+MT with adapters).
    Generalizations from one encoder-decoder and two speech+MT models to class-level conclusions are stated in Sections 3 and 7, which acknowledge that the choice is driven by performance and availability.
  • ad hoc to paper The attention-based probe is equally capable of extracting gender from pre-adapter and post-adapter representational spaces.
    The pre-ad vs. post-ad comparison in Section 4.1 is interpreted as adapter-induced erasure. If the probe is less compatible with the post-adapter embedding geometry, the drop could be an artifact rather than evidence of information removal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Different Speech Translation Models Encode and Translate Speaker Gender Differently." pith.science (2026). https://pith.science/paper/KP3LWCLE

@misc{pith2026250602172,
  author       = {Pith},
  title        = {Pith review of: Different Speech Translation Models Encode and Translate Speaker Gender Differently},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KP3LWCLE}},
  note         = {Machine review of arXiv:2506.02172}
}
read the original abstract

Recent studies on interpreting the hidden states of speech models have shown their ability to capture speaker-specific features, including gender. Does this finding also hold for speech translation (ST) models? If so, what are the implications for the speaker's gender assignment in translation? We address these questions from an interpretability perspective, using probing methods to assess gender encoding across diverse ST models. Results on three language directions (English-French/Italian/Spanish) indicate that while traditional encoder-decoder models capture gender information, newer architectures -- integrating a speech encoder with a machine translation system via adapters -- do not. We also demonstrate that low gender encoding capabilities result in systems' tendency toward a masculine default, a translation bias that is more pronounced in newer architectures.

Figures

Figures reproduced from arXiv: 2506.02172 by the authors.

Figure 1
Figure 1. Correlation between overall gender probing performance (macro F1) and gender translation accuracy across models and languages on test-speaker. higher translation quality increases the likelihood that system outputs include the speaker-referred words annotated in the reference. In all cases, more than half of the annotated words can be evaluated for gender translation. Regarding gender translation accuracy, enc-dec p… view at source ↗
Figure 2
Figure 2. Mean attention weights with standard deviations for the various attention-based probes. positions in the sequence are particularly impor￾tant for predicting gender, especially when the probe is highly effective. This aligns with pre￾vious work by Krishnan et al. (2024) and Attanasio et al. (2024), who found that gender encoding is predominantly concentrated at the beginning of the sequence. However, it is important … view at source ↗
Figure 3
Figure 3. Confusion matrices showing the relationship between [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Confusion matrices showing the relationship between [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 28 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Giuseppe Attanasio, Beatrice Savoldi, Dennis Fucci, and Dirk Hovy. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1188 Twists, humps, and pebbles: Multilingual speech recognition models exhibit gender performance gaps . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21318--21340, Miami, Florida, USA. As...

    doi:10.18653/v1/2024.emnlp-main.1188 2024

  4. [4]

    David Azul. 2013. How do voices become gendered? a critical examination of everyday and medical constructions of the relationship between voice, sex, and gender identity. In Challenging popular myths of sex, gender and biology, pages 77--88. Springer

    work page 2013

  5. [5]

    Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. https://arxiv.org/abs/2006.11477 wav2vec 2.0: A framework for self-supervised learning of speech representations . Preprint, arXiv:2006.11477

    arXiv 2020

  6. [6]

    Yonatan Belinkov. 2022. https://doi.org/10.1162/coli_a_00422 Probing classifiers: Promises, shortcomings, and advances . Computational Linguistics, 48(1):207--219

    doi:10.1162/coli_a_00422 2022

  7. [7]

    Yonatan Belinkov and James Glass. 2019. https://doi.org/10.1162/tacl_a_00254 Analysis methods in neural language processing: A survey . Transactions of the Association for Computational Linguistics, 7:49--72

    doi:10.1162/tacl_a_00254 2019

  8. [8]

    Luisa Bentivogli, Mauro Cettolo, Marco Gaido, Alina Karakanta, Alberto Martinelli, Matteo Negri, and Marco Turchi. 2021. https://doi.org/10.18653/v1/2021.acl-long.224 Cascade versus Direct Speech Translation: Do the Differences Still Make a Difference? In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th ...

    doi:10.18653/v1/2021.acl-long.224 2021

Show all 61 references
  1. [9]

    Di Gangi, Roldano Cattoni, and Marco Turchi

    Luisa Bentivogli, Beatrice Savoldi, Matteo Negri, Mattia A. Di Gangi, Roldano Cattoni, and Marco Turchi. 2020. https://doi.org/10.18653/v1/2020.acl-main.619 Gender in danger? evaluating speech translation technology on the M u ST - SHE corpus . In Proceedings of the 58th Annua...

    2020 doi

  2. [10]

    Su Lin Blodgett, Solon Barocas, Hal Daum \'e III, and Hanna Wallach. 2020. https://doi.org/10.18653/v1/2020.acl-main.485 Language (technology) is power: A critical survey of `` bias '' in NLP . In Proceedings of the 58th Annual Meeting of the Association for Computational Ling...

    2020 doi

  3. [11]

    Roldano Cattoni, Mattia Antonino Di Gangi , Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2021. https://doi.org/10.1016/j.csl.2020.101155 MuST-C: A multilingual corpus for end-to-end speech translation . Computer Speech & Language, 66:101155

    2021

  4. [12]

    Kwanghee Choi, Ankita Pasad, Tomohiko Nakamura, Satoru Fukayama, Karen Livescu, and Shinji Watanabe. 2024. https://doi.org/10.21437/Interspeech.2024-1157 Self-supervised speech representations are more phonetic than semantic . In Interspeech 2024, pages 4578--4582

    2024 doi

  5. [13]

    Shammur Absar Chowdhury, Nadir Durrani, and Ahmed Ali. 2024. https://doi.org/10.1016/j.csl.2023.101539 What do end-to-end speech models learn about speaker, language and channel information? a layer-wise and neuron-level analysis . Computer Speech & Language, 83:101539

    2024

  6. [14]

    Yu-An Chung, Yu Zhang, Wei Han, Chung-Cheng Chiu, James Qin, Ruoming Pang, and Yonghui Wu. 2021. https://arxiv.org/abs/2108.06209 W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training . Preprint, arXiv:2108.06209

    2021 arXiv

  7. [15]

    Ralph O. Coleman. 1976. https://doi.org/10.1044/jshr.1901.168 A comparison of the contributions of two voice quality characteristics to the perception of maleness and femaleness in the voice . Journal of Speech & Hearing Research, 19(1):168--180

    1976 doi

  8. [16]

    Seamless Communication, Loïc Barrault, Yu-An Chung, Mariano Cora Meglioli, David Dale, et al. 2023. https://arxiv.org/abs/2308.11596 SeamlessM4T: Massively Multilingual & Multimodal Machine Translation . Preprint, arXiv:2308.11596

    2023 arXiv

  9. [17]

    Alexis Conneau, German Kruszewski, Guillaume Lample, Lo \"i c Barrault, and Marco Baroni. 2018. https://doi.org/10.18653/v1/P18-1198 What you can cram into a single \ & ! \# * vector: Probing sentence embeddings for linguistic properties . In Proceedings of the 56th Annual Mee...

    2018 doi

  10. [18]

    Greville G. Corbett. 2013. The Expression of Gender. De Gruyter

    2013

  11. [19]

    Kelleher, and Julie Carson-Berndsen

    Patrick Cormac English, John D. Kelleher, and Julie Carson-Berndsen. 2022. https://doi.org/10.18653/v1/2022.sigmorphon-1.9 Domain-informed probing of wav2vec 2.0 embeddings for phonetic features . In Proceedings of the 19th SIGMORPHON Workshop on Computational Research in Phon...

    2022 doi

  12. [20]

    Maureen de Seyssel , Marvin Lavechin, Yossi Adi, Emmanuel Dupoux, and Guillaume Wisniewski. 2022. https://doi.org/10.21437/Interspeech.2022-373 Probing phoneme, language and speaker information in unsupervised speech representations . In Interspeech 2022, pages 1402--1406

    2022 doi

  13. [21]

    Yanai Elazar, Shauli Ravfogel, Alon Jacovi, and Yoav Goldberg. 2021. https://doi.org/10.1162/tacl_a_00359 Amnesic probing: Behavioral explanation with amnesic counterfactuals . Transactions of the Association for Computational Linguistics, 9:160--175

    2021 doi

  14. [22]

    Costa-jussà

    Javier Ferrando, Gabriele Sarti, Arianna Bisazza, and Marta R. Costa-jussà. 2024. https://arxiv.org/abs/2405.00208 A Primer on the Inner Workings of Transformer-based Language Models . Preprint, arXiv:2405.00208

    2024 arXiv

  15. [23]

    Charlie Frogner, Chiyuan Zhang, Hossein Mobahi, Mauricio Araya-Polo, and Tomaso Poggio. 2015. Learning with a wasserstein loss. In Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 2, NIPS'15, page 2053–2061, Cambridge, MA, USA....

    2015

  16. [24]

    Susanne Fuchs and Martine Toda. 2010. Do differences in male versus female/s/reflect biological or sociophonetic factors. Turbulent sounds: An interdisciplinary guide, 21:281--302

    2010

  17. [25]

    Marco Gaido, Sara Papi, Matteo Negri, and Luisa Bentivogli. 2024. https://arxiv.org/abs/2402.12025 Speech Translation with Speech Foundation Models and Large Language Models: What is There and What is Missing? Preprint, arXiv:2402.12025

    2024 arXiv

  18. [26]

    Marco Gaido, Beatrice Savoldi, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2020. https://doi.org/10.18653/v1/2020.coling-main.350 Breeding gender-aware direct speech translation systems . In Proceedings of the 28th International Conference on Computational Linguistics, p...

    2020 doi

  19. [27]

    Xavier Glorot and Yoshua Bengio. 2010. https://proceedings.mlr.press/v9/glorot10a.html Understanding the difficulty of training deep feedforward neural networks . In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, volume 9 of P...

    2010

  20. [28]

    Séverine Guillaume, Maxime Fily, Alexis Michaud, and Guillaume Wisniewski. 2024. https://doi.org/10.21437/Interspeech.2024-953 Gender and language identification in multilingual models of speech: Exploring the genericity and robustness of speech representations . In Interspeec...

    2024 doi

  21. [29]

    Gygax, Daniel Elmiger, Sandrine Zufferey, Alan Garnham, Sabine Sczesny, Friederike von Stockhausen, Lisa Braun, and Jane Oakhill

    Pascal M. Gygax, Daniel Elmiger, Sandrine Zufferey, Alan Garnham, Sabine Sczesny, Friederike von Stockhausen, Lisa Braun, and Jane Oakhill. 2019. https://doi.org/10.3389/fpsyg.2019.01604 A language index of grammatical gender dimensions to study the impact of grammatical gende...

    2019

  22. [30]

    John Hewitt and Percy Liang. 2019. https://doi.org/10.18653/v1/D19-1275 Designing and interpreting probes with control tasks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Langua...

    2019 doi

  23. [31]

    Hillenbrand and Michael J

    James M. Hillenbrand and Michael J. Clark. 2009. https://doi.org/10.3758/APP.71.5.1150 The role of f0 and formant frequencies in distinguishing the voices of men and women . Attention Perception & Psychophysics, 71(5):1150--1166

    2009 doi

  24. [32]

    Abdullah, and Dietrich Klakow

    Aravind Krishnan, Badr M. Abdullah, and Dietrich Klakow. 2024. https://doi.org/10.21437/Interspeech.2024-2209 On the encoding of gender in transformer-based asr representations . In Interspeech 2024, pages 3090--3094

    2024 doi

  25. [33]

    Yeptain Leung, Jennifer Oates, and Siew Pang Chan. 2018. Voice, articulation, and prosody contribute to listener perceptions of speaker gender: A systematic review and meta-analysis. Journal of Speech, Language, and Hearing Research, 61(2):266--297

    2018

  26. [34]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org

    2023

  27. [35]

    Nayla Matar, Cristel Portes, Leonardo Lancia, Thierry Legou, and Fabienne Baider. 2016. https://hal.archives-ouvertes.fr/hal-01459619 Voice quality and gender stereotypes: A study on Lebanese women with Reinke's edema . Journal of Speech, Language, and Hearing Research , 59(6)...

    2016

  28. [36]

    Danielle Pereira Menezes, Zulina Souza de Lira , Ana Nery Barbosa de Araújo, Anna Alice Figueirêdo de Almeida , Adriana de Oliveira Camargo Gomes, Bruno Teixeira Moraes, and Jonia Alves Lucena. 2022. https://doi.org/10.1016/j.jvoice.2021.12.020 Prosodic differences in the voic...

    2022 doi

  29. [37]

    Frida Nylén, Johan Holmberg, and Maria Södersten. 2024. https://doi.org/10.1121/10.0025932 Acoustic cues to femininity and masculinity in spontaneous speech . The Journal of the Acoustical Society of America, 155(5):3090--3100

    2024 doi

  30. [38]

    Ankita Pasad, Bowen Shi, and Karen Livescu. 2023. https://doi.org/10.1109/ICASSP49357.2023.10096149 Comparative layer-wise analysis of self-supervised speech models . In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1--5

    2023

  31. [39]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K\" o pf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu...

    2019

  32. [40]

    Fabian Pedregosa, Ga\" e l Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and \' E douard ...

    2011

  33. [41]

    Amanda Maria Pereira, Ana Paula Dassie-Leite, Eliane Cristina Pereira, Juliana Benthien Cavichiolo, Marcelo de Oliveira Rosa, and Elmar Allen Fugmann. 2018. https://doi.org/10.1590/2317-1782/20182017046 Auditory perception of lay judges about gender identification of women wit...

    2018 doi

  34. [42]

    Gabriel Peyré and Marco Cuturi. 2019. https://doi.org/10.1561/2200000073 Computational optimal transport: With applications to data science . Foundations and Trends® in Machine Learning, 11(5-6):355--607

    2019 doi

  35. [43]

    Archiki Prasad and Preethi Jyothi. 2020. https://doi.org/10.18653/v1/2020.acl-main.345 How accents confound: Probing for accent information in end-to-end speech recognition systems . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pa...

    2020 doi

  36. [44]

    Abhilasha Ravichander, Yonatan Belinkov, and Eduard Hovy. 2021. https://doi.org/10.18653/v1/2021.eacl-main.295 Probing the probing paradigm: Does probing accuracy entail task relevance? In Proceedings of the 16th Conference of the European Chapter of the Association for Comput...

    2021 doi

  37. [45]

    Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.213 COMET : A neural framework for MT evaluation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685--2702, ...

    2020 doi

  38. [46]

    Naomi Saphra and Sarah Wiegreffe. 2024. https://doi.org/10.18653/v1/2024.blackboxnlp-1.30 Mechanistic? In Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 480--498, Miami, Florida, US. Association for Computational Linguistics

    2024 doi

  39. [47]

    Beatrice Savoldi, Marco Gaido, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2021. https://doi.org/10.1162/tacl_a_00401 Gender bias in machine translation . Transactions of the Association for Computational Linguistics, 9:845--874

    2021 doi

  40. [48]

    Beatrice Savoldi, Marco Gaido, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2022. https://doi.org/10.18653/v1/2022.acl-long.127 Under the morphosyntactic lens: A multifaceted evaluation of gender bias in speech translation . In Proceedings of the 60th Annual Meeting of th...

    2022 doi

  41. [49]

    Gaofei Shen, Michaela Watkins, Afra Alishahi, Arianna Bisazza, and Grzegorz Chrupa a. 2024. https://doi.org/10.18653/v1/2024.naacl-long.239 Encoding of lexical tone in self-supervised models of spoken language . In Proceedings of the 2024 Conference of the North American Chapt...

    2024 doi

  42. [50]

    Adrian P. Simpson. 2001. https://doi.org/10.1121/1.1356020 Dynamic consequences of differences in male and female vocal tract dimensions . The Journal of the Acoustical Society of America, 109(5 Pt 1):2153--2164

    2001 doi

  43. [51]

    Matthias Sperber and Matthias Paulik. 2020. https://doi.org/10.18653/v1/2020.acl-main.661 Speech translation and the end-to-end promise: Taking stock of where we are . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7409--7421,...

    2020 doi

  44. [52]

    Dagmar Stahlberg, Friederike Braun, Lisa Irmen, and Sabine Sczesny. 2007. https://psycnet.apa.org/record/2007-01308-006 Representation of the sexes in language . In Klaus Fiedler, editor, Social Communication, pages 163--187. Psychology Press

    2007

  45. [53]

    Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2019. https://doi.org/10.18653/v1/P19-1159 Mitigating gender bias in natural language processing: Literature review . In Proceeding...

    2019 doi

  46. [54]

    Costa-jussà, James Cross, Onur Çelebi, et al

    NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, et al. 2022. https://arxiv.org/abs/2207.04672 No language left behind: Scaling human-centered machine translation . Preprint, arXiv:2207.04672

    2022 arXiv

  47. [55]

    Ioannis Tsiamas, Gerard G \'a llego, Jos \'e Fonollosa, and Marta Costa-juss \`a . 2024. https://doi.org/10.18653/v1/2024.findings-acl.847 Pushing the limits of zero-shot end-to-end speech translation . In Findings of the Association for Computational Linguistics: ACL 2024, pa...

    2024 doi

  48. [56]

    Anna Paula Villas-Bôas, Karine Schwarz, Anna Martha Vaitses Fontanari, Angelo Brandelli Costa, Dhiordan Cardoso da Silva, Maiko Abel Schneider, Carla Aparecida Cielo, Poli Mara Spritzer, and Maria Inês Rodrigues Lobato. 2021. https://doi.org/10.3389/fpsyg.2021.622526 Acoustic ...

    2021

  49. [57]

    Abdul Waheed, Hanin Atwany, Bhiksha Raj, and Rita Singh. 2024. https://arxiv.org/abs/2410.12948 What do speech foundation models not learn about speech? Preprint, arXiv:2410.12948

    2024 arXiv

  50. [58]

    Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, and Juan Pino. 2020. fairseq s2t: Fast speech-to-text modeling with fairseq. In Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations

    2020

  51. [59]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020 doi

  52. [60]

    Mu Yang, Ram C. M. C. Shekar, Okim Kang, and John H. L. Hansen. 2023. https://doi.org/10.21437/Interspeech.2023-2254 What can an accent identifier learn? probing phonetic and prosodic information in a wav2vec2-based accent identification model . In Interspeech 2023, pages 1923--1927

    2023 doi

  53. [61]

    Jinming Zhao, Hao Yang, Gholamreza Haffari, and Ehsan Shareghi. 2022. https://doi.org/10.21437/Interspeech.2022-592 M-adapter: Modality adaptation for end-to-end speech-to-text translation . In Interspeech 2022, pages 111--115

    2022 doi

Pith tools

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