Pith. sign in

REVIEW 5 major objections 5 minor 27 references

Multimodal Zero-Shot Framework for Deepfake Hate Speech Detection in Low-Resource Languages

T0 review · 5 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 The paper claims that a contrastive audio-text alignment detector, trained on paired transcripts and synthesized speech in English and five low-resource Indian languages, outperforms text-only and multimodal baselines in both in-domain…

desk verdict The dataset is a real resource, but the reported accuracies are below a majority-class baseline, so the main performance and zero-shot claims do not hold as written. read the letter →

arxiv 2506.08372 v1 pith:LMQMGOGG submitted 2025-06-10 cs.SD eess.AS

Rishabh Ranjan , Likhith Ayinala , Mayank Vatsa , Richa Singh This is my paper
classification cs.SDeess.AS
keywords hatespeechdetectionaudiodeepfakesmultimodallearningcontrastivezero-shotclassificationlow-resourcelanguagessyntheticmultilingual
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

The paper sets out to show that hate speech in synthesized (deepfake-style) audio is detected more reliably when the audio is judged together with its text transcription, with both modalities mapped into one shared embedding space by contrastive learning. To support the claim, the authors build and release a benchmark of 127,290 paired text-and-speech samples in English, Hindi, Bengali, Marathi, Tamil, and Telugu, generated by reading existing text corpora aloud with a multilingual text-to-speech system. On in-domain splits their method reaches 0.819 and 0.701 accuracy on two language sets, and on cross-language zero-shot splits it reaches 0.625 and 0.786, ahead of every baseline in all four settings. The practical point is that content moderation increasingly faces manipulated audio in low-resource languages, where text-only detectors have little training data to rely on.

What carries the argument

The load-bearing mechanism is a shared semantic embedding space produced by contrastive alignment between a SeamlessM4T audio encoder and a SONAR text encoder, two multilingual pretrained networks. Audio and text embeddings are normalized onto a unit hypersphere, and a batch-wise similarity matrix with positive and negative masks based on labels drives same-label audio-text pairs together and different-label pairs apart. In the downstream phase, concatenated embeddings feed a classifier trained with a weighted sum of triplet loss and binary cross-entropy loss, where triplets are formed across languages so the hate/non-hate distinction is forced to become language-independent. This common vector space is what lets a classifier trained on one set of languages score audio-text pairs from another set without retraining.

What would settle it

Run the same SONAR and SeamlessM4T encoders on the same splits with plain concatenation and cross-entropy loss, omitting the contrastive alignment; if accuracy matches the reported 0.819/0.701 and zero-shot numbers, alignment is not what carries the gain. Alternatively, test on a language verifiably absent from both encoders' pretraining data; if accuracy falls to chance, the zero-shot result is best explained by pretraining coverage rather than by the proposed mechanism.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that a two-stage contrastive procedure creates a language-invariant hate-speech classifier for synthetic speech. The audio encoder is initialized from SeamlessM4T and the text encoder from SONAR; embeddings are L2-normalized, paired audio-text similarities are computed in batches, and a label-based positive/negative mask pulls same-label pairs together while pushing different-label pairs apart. A downstream classifier concatenates the two embeddings and is trained with triplet loss plus binary cross-entropy loss. The method reports accuracy 0.819 on Set-A (Marathi, Bengali, Tamil) and 0.701 on Set-B (English, Hindi, Telugu), outperforming the strongest baseline WAVELMBERT at 0.790 and 0.669; in the cross-subset zero-shot protocol it reports 0.625 when trained on Set-A and evaluated on Set-B, and 0.786 in the reverse direction. The authors interpret the consistent margin over text-only models as evidence that acoustic features carry hate-speech signal beyond the transcript.

Load-bearing premise

The load-bearing premise is that the evaluation languages are genuinely new to the pretrained models, so the cross-language accuracy can be credited to the proposed alignment; in fact, the audio and text models were trained on many languages, including the ones used for testing.

Editorial extensions

If this is right

  • Adding a paired text channel to an audio hate-speech detector should improve accuracy on synthesized speech, since the paper's multimodal models beat all text-only baselines on every reported split.
  • Cross-language transfer is asymmetric: training on the set that includes English transfers better to the low-resource set (0.786) than training on the low-resource set transfers back (0.625), suggesting that exposure to English helps the shared embedding space generalize.
  • A language can be added to the detector without retraining the classifier, provided its audio and text pass through the same encoders into the shared space, which is the operational meaning of zero-shot in the paper.
  • The released benchmark gives other groups a direct way to test whether the reported superiority survives with different audio encoders, TTS voices, or annotation schemes.

Reading between the lines

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

  • Editorial inference: because SONAR and SeamlessM4T are pretrained on many languages beyond the six studied, the zero-shot numbers are best read as generalization to languages unseen at fine-tuning time, not as evidence about truly novel languages; the reported gains may partly reflect the encoders' existing multilingual coverage.
  • Editorial inference: the dataset is synthesized speech read from text corpora by a single TTS system, not real manipulated audio, so the method has been tested on deepfake-like prosody rather than on tampered recordings; extending the same contrastive alignment to natural spoofed audio and multiple speakers is the immediate next experiment.
  • Reproducibility note in the manuscript: reference [13], cited as the multilingual TTS model that produced the audio, points to an unrelated computer-science textbook, and no citation is given for the SONAR text encoder; correcting these provenance gaps is needed before the dataset and encoder setup can be independently reconstructed.
  • Editorial inference: a decisive ablation would compare the full contrastive pipeline against the same two encoders with simple concatenation and cross-entropy loss; the paper does not report it, and without it the specific contribution of contrastive pretraining remains untested.
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

5 major / 5 minor

Summary. The paper proposes a multimodal framework for hate speech detection in synthesized (TTS) audio, using SONAR for text and SeamlessM4T for audio, with a two-stage training procedure: a contrastive learning stage that aligns audio and text embeddings followed by a downstream stage combining triplet loss and binary cross-entropy. It also introduces a multilingual dataset with paired text and TTS audio across English and five Indian languages, and evaluates the approach on two language sets (Set-A and Set-B) plus cross-set zero-shot settings, reporting accuracy gains over six baselines. The central claims are that the method outperforms text-only and multimodal baselines and generalizes to unseen languages.

Significance. If the evaluation were methodologically sound, the dataset and the contrastive alignment idea would be a useful contribution: the dataset addresses a gap in speech-based hate speech detection in low-resource languages, and the cross-modal contrastive approach is a reasonable direction. However, the reported evaluation contains internal inconsistencies and uses a threshold-dependent metric on highly imbalanced test sets, so as presented the paper does not establish the claimed superiority or zero-shot generalization.

major comments (5)
  1. [Table 1 and §3.1] The test sets shown in Table 1 are so class-imbalanced that a trivial classifier that always predicts 'hate' achieves higher accuracy than the proposed model. For Set-A, the combined test set contains 30,831 hate and 3,266 non-hate samples (90.4% hate), so a constant 'hate' predictor has accuracy 0.904, exceeding the reported 0.819; for Set-B, the test set is 35,905 hate and 11,831 non-hate (75.2% hate), so the trivial accuracy is 0.752, exceeding the reported 0.701. The same issue applies to the cross-set results in Table 3 (0.625 vs 0.752 and 0.786 vs 0.904). Thus the reported ACC values cannot support the claim of outperformance; the authors must report threshold-independent metrics such as AUC-ROC, balanced accuracy, or F1, and should reconcile these numbers with the claimed 70:30 split.
  2. [§3.1 and Table 1] The stated 70:30 train/test split is inconsistent with Table 1: Telugu is 98.9% test (34,772 of 35,142 samples), Marathi is 50:50, and English is about 57% test. In addition, the Bengali row is internally inconsistent: 7,000 + 14,000 + 3,000 + 600 = 24,600, yet the table lists Total as 30,000; the grand total of 127,290 does not match the sum of the printed rows (121,890 if Bengali is 24,600). These discrepancies call into question the dataset statistics and the correctness of the evaluation protocol, and they must be corrected or explained.
  3. [§3.2 and Tables 2–3] Section 3.2 states that performance is evaluated using accuracy, F1-score, and AUC-ROC, but Tables 2 and 3 report only ACC and EER. The missing metrics are particularly important given the class imbalance documented in Table 1. Moreover, no error bars, confidence intervals, or multiple-run variability are reported, so it is impossible to determine whether the differences between the proposed model and the baselines are statistically meaningful.
  4. [§2.1 and §3.1] The zero-shot claim is confounded by the selection of pretrained encoders: SONAR and SeamlessM4T are massively multilingual models that were pretrained on many languages, very likely including the evaluation languages in Set-A and Set-B. Consequently, the cross-set experiments do not test generalization to languages entirely unseen by the encoders; they only test the classifier head on languages the encoders already cover. The paper overstates the zero-shot nature and should either use encoders that genuinely exclude the test languages, or clearly reframe the claim as 'unseen at the classifier level' and note the pretraining overlap.
  5. [§2 and Table 2] The comparison with baselines is not controlled: the proposed model uses SONAR and SeamlessM4T, while the multimodal baselines use HUBERT, WavLM, or AST combined with BERT, and the text baselines use a different BERT tokenizer. Since the proposed model has access to different (and potentially stronger) pretrained representations, the observed accuracy gains cannot be attributed specifically to the contrastive alignment or the two-stage loss. An ablation that keeps the same encoders across the proposed model and a baseline classification head (or that uses the same backbone for the proposed and baselines) is necessary to support the claimed advantage.
minor comments (5)
  1. [§2.1] The contrastive loss L is described only in words; the paper never gives its explicit mathematical form, which makes the method incompletely specified.
  2. [Eq. (1)] The notation in Eq. (1) introduces s_u for the output but the downstream phase later uses \hat{y}_i; the symbol s_u is never defined and appears to be a typographical leftover.
  3. [References] Reference [13] is cited as the source of the Massive Multi-Lingual TTS model, but the reference is to 'Structure and Interpretation of Computer Programs' (Abelson, Sussman, Sussman), which is unrelated; the correct source for Meta's Massive Multilingual Speech model should be cited.
  4. [Title and Abstract] The term 'deepfake' is used loosely: the audio is generated by TTS, which is synthetic speech but not necessarily a deepfake impersonation of a specific person. The paper should clarify what is meant by deepfake audio in this context.
  5. [§3.1] The description '70:30 ratio' is contradicted by Table 1 for every language, as documented in the major comments; if the intended split was something else (e.g., column-wise pre-allocated ratios per source corpus), the text should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the reported evaluations are held-out empirical comparisons, and the zero-shot leakage and Table 1 arithmetic issues are correctness risks rather than circular reductions.

full rationale

The paper's claimed derivations are all supervised or held-out empirical evaluations. Section 2.1 defines a contrastive loss whose positive and negative masks are built from labels, and Section 2.2 fine-tunes with triplet plus binary cross-entropy loss; the test numbers in Tables 2 and 3 come from separate train/test partitions, so no reported accuracy is equal by construction to a fitted parameter or label statistic. The encoders are initialized from external SONAR and SeamlessM4T weights (Section 2.1), not from the authors' own prior results, and no load-bearing claim rests on a self-citation or on an imported uniqueness theorem. The 'zero-shot' claim is confounded: Section 1 defines zero-shot as 'detecting hate speech in new languages,' while Section 2.1 relies on SONAR and SeamlessM4T, which are pretrained on many languages including the evaluation languages, so cross-subset accuracy may partly reflect pretraining coverage rather than the proposed alignment. That is a validity and threat-to-inference issue, not a case where the predicted quantity reduces to its input by definition. Similarly, Section 3.1 states a 70:30 split, but Table 1's per-language rows (e.g., Telugu: 259+111 train versus 24,340+10,432 test) contradict that split, and the printed totals do not add to the claimed 127,290; if the test sets are as imbalanced as the table suggests, the reported accuracies may fall below a constant-classifier baseline. These table issues undermine the empirical claims but do not constitute circular reasoning in the derivation sense. No step in the paper, on the evidence quoted, defines a predicted value in terms of the target output or fits a parameter and then relabels the resulting value as a prediction.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claims rest on external pretrained encoders and on the assumption that synthetic audio inherits text labels. Both are plausible but unverified, and the zero-shot claim additionally assumes that the encoders have not already seen the evaluation languages.

free parameters (3)
  • alpha (loss weight)
    Weight balancing triplet loss and binary cross-entropy in Eq. 3; no value is reported, so the training objective is not fully specified.
  • triplet margin
    Margin for the triplet loss in the downstream phase; not mentioned in the paper, though it directly affects the embedding geometry.
  • embedding dimension m
    Output dimension of the shared embedding space V; the paper writes V subset R^{m x d} but never states the dimension used in experiments.
assumptions (3)
  • domain assumption Labels of source text remain valid for synthesized audio.
    Section 3 converts existing text corpora to TTS audio and reuses the original labels without verifying that the spoken rendering preserves the hateful/non-hateful content.
  • domain assumption SONAR and SeamlessM4T embeddings can be mapped into a common semantic space by contrastive alignment.
    Section 2.1 initializes the encoders from these pretrained models and assumes their outputs are comparable; the paper provides no evidence that the two embedding spaces are initially aligned.
  • domain assumption A classifier trained on a subset of languages generalizes to the other languages (zero-shot).
    Section 3.1 defines cross-subset protocols, but the pretrained encoders already cover the evaluation languages, so the zero-shot premise is not independently established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Zero-Shot Framework for Deepfake Hate Speech Detection in Low-Resource Languages." pith.science (2026). https://pith.science/paper/LMQMGOGG

@misc{pith2026250608372,
  author       = {Pith},
  title        = {Pith review of: Multimodal Zero-Shot Framework for Deepfake Hate Speech Detection in Low-Resource Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LMQMGOGG}},
  note         = {Machine review of arXiv:2506.08372}
}
read the original abstract

This paper introduces a novel multimodal framework for hate speech detection in deepfake audio, excelling even in zero-shot scenarios. Unlike previous approaches, our method uses contrastive learning to jointly align audio and text representations across languages. We present the first benchmark dataset with 127,290 paired text and synthesized speech samples in six languages: English and five low-resource Indian languages (Hindi, Bengali, Marathi, Tamil, Telugu). Our model learns a shared semantic embedding space, enabling robust cross-lingual and cross-modal classification. Experiments on two multilingual test sets show our approach outperforms baselines, achieving accuracies of 0.819 and 0.701, and generalizes well to unseen languages. This demonstrates the advantage of combining modalities for hate speech detection in synthetic media, especially in low-resource settings where unimodal models falter. The Dataset is available at https://www.iab-rubric.org/resources.

Figures

Figures reproduced from arXiv: 2506.08372 by the authors.

Figure 1
Figure 1. (A) Current hate speech detectors often misclassify deepfake-generated hate speech audio as non-hateful, exposing moderation systems to manipulation. (B) The proposed multi￾modal framework effectively distinguishes deepfake hate speech from genuine non-hateful content, enhancing detection accu￾racy and robustness. multilingual audio-based toxicity detectors capable of zero-shot detection across multiple languages, r… view at source ↗
Figure 2
Figure 2. Proposed multimodal hate speech detection pipeline: Audio and text inputs are encoded separately and mapped into a common semantic embedding space, enabling effec￾tive cross-modal and cross-lingual classification through con￾trastive learning. modal dataset addressing the gap in deepfake audio hate speech detection. Second, we introduce a novel, contrastive learning￾based architecture that effectively fuses textual … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 22 canonical work pages

  1. [1]

    Multimodal Zero-Shot Framework for Deepfake Hate Speech Detection in Low-Resource Languages

    Introduction Social media has transformed global communication, connect- ing nearly 6.3 billion users and exhibiting a remarkable com- pound annual growth rate-driven primarily by emerging mar- kets in Asia, such as India, China, and Indonesia. However, this digital expansion has also amplified the reach of hate speech. Online hate speech demonstrably har...

    work page Pith review arXiv 2025

  2. [2]

    Proposed Hate Audio Detector Model The proposed framework introduces a novel two-stage con- trastive learning approach for multimodal hate speech detection, distinctly combining audio and text modalities. Unlike existing methods, we employ a unique alignment strategy in the pre- training stage, leveraging state-of-the-art encoders, SONAR for text and Seam...

    work page 1926

  3. [3]

    Proposed Dataset and Experimental Details We introduce a novel multimodal, multilingual dataset designed to address the scarcity of resources for hate speech detection, with a focus on low-resource languages. The dataset encom- passes both textual and audio modalities across six languages: English, Hindi, Telugu, Tamil, Marathi, and Bengali, chosen for th...

    work page

  4. [4]

    These experiments address RO1 by demon- strating the effectiveness of our approach for multilingual deep- fake hate speech detection across diverse language groups

    Results and Analysis RO1: Multilingual Deepfake Hate Speech Detection.We evaluate the performance of our proposed model on two lan- guage sets: Set-A (Tamil, Marathi, Bengali) and Set-B (English, Hindi, Telugu). These experiments address RO1 by demon- strating the effectiveness of our approach for multilingual deep- fake hate speech detection across diver...

    work page

  5. [5]

    Conclusion This paper presents a novel zero-shot hate speech detection ap- proach for audio, addressing challenges posed by multimodal content in low-resource languages. Our contributions include: (1) a framework that integrates audio and text modalities us- ing contrastive learning; (2) the first comprehensive multimodal dataset for deepfake hate speech ...

    work page

  6. [6]

    The authors also gratefully acknowledge the support of IndiaAI and Meta through Srijan: Centre of Excellence for Generative AI

    Acknowledgement This research is supported by a grant from the NSM, MeitY . The authors also gratefully acknowledge the support of IndiaAI and Meta through Srijan: Centre of Excellence for Generative AI

    work page

  7. [7]

    Fanning the flames of hate: So- cial media and hate crime,

    K. M ¨uller and C. Schwarz, “Fanning the flames of hate: So- cial media and hate crime,” June 5 2020, available at SSRN: https://ssrn.com/abstract=3082972 or http://dx.doi.org/10.2139/ ssrn.3082972

    work page 2020

  8. [8]

    Hatexplain: A benchmark dataset for explainable hate speech detection,

    B. Mathew, P. Saha, S. M. Yimam, C. Biemann, P. Goyal, and A. Mukherjee, “Hatexplain: A benchmark dataset for explainable hate speech detection,” inAAAI, 2021, pp. 14 867–14 875

    work page 2021

Show all 27 references
  1. [9]

    Multilingual and multi-aspect hate speech analysis,

    N. Ousidhoum, Z. Lin, H. Zhang, Y . Song, and D.-Y . Yeung, “Multilingual and multi-aspect hate speech analysis,” inProc. of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP...

    2019

  2. [10]

    Multilingual detection of hate speech against im- migrants and women in twitter at semeval-2019 task 5: Fre- quency analysis interpolation for hate in speech detection,

    `O. G. i Orts, “Multilingual detection of hate speech against im- migrants and women in twitter at semeval-2019 task 5: Fre- quency analysis interpolation for hate in speech detection,” in SemEval@NAACL-HLT. Association for Computational Lin- guistics, 2019, pp. 460–463

    2019

  3. [11]

    Ethos: a multi-label hate speech detection dataset,

    I. Mollas, Z. Chrysopoulou, S. Karlos, and G. Tsoumakas, “Ethos: a multi-label hate speech detection dataset,”Complex & Intelli- gent Systems, vol. 8, no. 6, pp. 4663–4678, 2022

    2022

  4. [12]

    Seamlessm4t-massively multilingual & multimodal machine translation,

    S. Communication, L. Barrault, Y . Chung, M. C. Meglioli, D. Dale, N. Dong, P. Duquenne, H. Elsahar, H. Gong, K. Hef- fernan, J. Hoffman, C. Klaiber, P. Li, D. Licht, J. Maillard, A. Rakotoarison, K. R. Sadagopan, G. Wenzek, E. Ye, B. Akula, P. Chen, N. E. Hachem, B. Ellis, G....

    2023 arXiv

  5. [13]

    Detoxy: A large-scale multimodal dataset for toxicity classifi- cation in spoken utterances,

    S. Ghosh, S. Lepcha, S. Singh, R. R. Shah, and S. Umesh, “Detoxy: A large-scale multimodal dataset for toxicity classifi- cation in spoken utterances,” inINTERSPEECH. ISCA, 2022, pp. 5185–5189

    2022

  6. [14]

    Audio-based toxic language classification using self-attentive convolutional neural network,

    M. Yousefi and D. Emmanouilidou, “Audio-based toxic language classification using self-attentive convolutional neural network,” in29th European Signal Processing Conference, EUSIPCO 2021, Dublin, Ireland, August 23-27, 2021. IEEE, 2021, pp. 11–15. [Online]. Available: https://...

    2021

  7. [15]

    Mutox: Universal multilingual audio-based toxicity dataset and zero-shot detector,

    M. R. Costa-juss `a, M. C. Meglioli, P. Andrews, D. Dale, P. Hansanti, E. Kalbassi, A. Mourachko, C. Ropers, and C. Wood, “Mutox: Universal multilingual audio-based toxicity dataset and zero-shot detector,” inFindings of the Association for Computational Linguistics, ACL 2024,...

    2024

  8. [16]

    Hostility detection dataset in hindi,

    M. Bhardwaj, M. S. Akhtar, A. Ekbal, A. Das, and T. Chakraborty, “Hostility detection dataset in hindi,”ArXiv, vol. abs/2011.03588, 2020. [Online]. Available: https://api. semanticscholar.org/CorpusID:226281866

    2011 arXiv

  9. [17]

    Am I a resource-poor language? data sets, embed- dings, models and analysis for four different NLP tasks in telugu language,

    M. Marreddy, S. R. Oota, L. S. Vakada, V . C. Chinni, and R. Mamidi, “Am I a resource-poor language? data sets, embed- dings, models and analysis for four different NLP tasks in telugu language,”ACM Trans. Asian Low Resour. Lang. Inf. Process., vol. 22, no. 1, pp. 18:1–18:34, 2023

    2023

  10. [18]

    Adima: Abuse detection in multilingual audio,

    V . Gupta, R. Sharon, R. Sawhney, and D. Mukherjee, “Adima: Abuse detection in multilingual audio,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6172–6176

    2022

  11. [19]

    Abelson, G

    H. Abelson, G. J. Sussman, and J. Sussman,Structure and Inter- pretation of Computer Programs. Cambridge, Massachusetts: MIT Press, 1985

    1985

  12. [20]

    L3cube-mahahate: A tweet- based marathi hate speech detection dataset and bert models,

    H. Patil, A. Velankar, and R. Joshi, “L3cube-mahahate: A tweet- based marathi hate speech detection dataset and bert models,” in Proceedings of the Third Workshop on Threat, Aggression and Cyberbullying (TRAC 2022), 2022, pp. 1–9

    2022

  13. [21]

    Hate speech detection in the bengali language: A dataset and its baseline evaluation,

    N. Romim, M. F. Ahmed, H. Talukder, and M. S. Islam, “Hate speech detection in the bengali language: A dataset and its baseline evaluation,”ArXiv, vol. abs/2012.09686, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:229298046

    2012 arXiv

  14. [22]

    Hate speech, machine classification and statistical modelling of information flows on twitter: interpretation and communication for policy decision making,

    P. Burnap and M. L. Williams, “Hate speech, machine classification and statistical modelling of information flows on twitter: interpretation and communication for policy decision making,” 2014. [Online]. Available: https://api.semanticscholar. org/CorpusID:142840744

    2014

  15. [23]

    Overview of the hasoc subtrack at fire 2022: Hate speech and offensive content identification in english and indo-aryan languages,

    S. Satapara, P. Majumder, T. Mandl, S. Modha, H. Madhu, T. Ranasinghe, M. Zampieri, K. North, and D. Premasiri, “Overview of the hasoc subtrack at fire 2022: Hate speech and offensive content identification in english and indo-aryan languages,” inProceedings of the 14th Annual...

    2022 doi

  16. [24]

    Detecting hate speech on twitter using a convolution-gru based deep neural network,

    Z. Zhang, D. Robinson, and J. A. Tepper, “Detecting hate speech on twitter using a convolution-gru based deep neural network,” inESWC, ser. Lecture Notes in Computer Science, vol. 10843. Springer, 2018, pp. 745–760

    2018

  17. [25]

    AST: audio spectrogram transformer,

    Y . Gong, Y . Chung, and J. R. Glass, “AST: audio spectrogram transformer,”CoRR, vol. abs/2104.01778, 2021. [Online]. Available: https://arxiv.org/abs/2104.01778

    2021 arXiv

  18. [26]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units,

    W. Hsu, B. Bolte, Y . H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,”CoRR, vol. abs/2106.07447, 2021. [Online]. Available: https://arxiv.org/abs/ 2106.07447

    2021 arXiv

  19. [27]

    Wavlm: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, and F. Wei, “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,”CoRR, vol. abs/2110.13900, 2021...

Pith tools

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