Towards Cross-Lingual Audio Abuse Detection in Low-Resource Settings with Few-Shot Learning

Aditya Narayan Sankaran; Noel Crespi; Reza Farahbakhsh

REVIEW 4 major objections 6 minor 56 references

Towards Cross-Lingual Audio Abuse Detection in Low-Resource Settings with Few-Shot Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper argues that a MAML-based few-shot classifier on pre-trained audio features can detect abusive speech in ten Indian languages with only 50–200 labeled clips per language, with Whisper's $L_2$-normalized embeddings reaching…

desk verdict Useful empirical results, but the paper's central MAML few-shot claim is not backed by the written protocol. read the letter →

arxiv 2412.01408 v3 pith:XKID2JVN submitted 2024-12-02 cs.CL cs.AI

Aditya Narayan Sankaran , Reza Farahbakhsh , Noel Crespi This is my paper · ORCID

classification cs.CLcs.AI

keywords few-shotlearningaudioabusedetectioncross-lingualtransferlow-resourcelanguagesModel-AgnosticMeta-Learning(MAML)WhisperADIMAdatasetfeaturenormalization

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

This paper argues that few-shot meta-learning on top of pre-trained audio features can make audio abuse detection practical for low-resource languages, using ten Indian languages as the test bed. The authors combine the MAML meta-learning algorithm with fixed embeddings from Whisper and Wav2Vec, evaluating two feature-normalization schemes at 50, 100, 150, and 200 labeled samples per language. Whisper features normalized with the $L_2$-norm reach 78.98%–85.22% accuracy in the 100-shot setting, and the best macro-F1 scores beat or approach the zero-shot baseline from the ADIMA dataset in most languages with reported numbers. The paper also presents a t-SNE feature study suggesting that acoustic similarity between related languages supports cross-lingual transfer.

What carries the argument

The load-bearing mechanism is MAML, a meta-learning algorithm that trains a model to adapt quickly to a new task from a few examples, applied to fixed pre-trained audio features. The features come from Whisper-large and the CLSRIL-23 Wav2Vec variant, pooled by temporal-mean or $L_2$-norm normalization, and a three-layer fully connected network is the learner. The paper samples $k$ clips per language for support and trains across all ten languages, varying $k$ from 50 to 200.

What would settle it

Train a three-layer MLP with the same architecture, features, and merged support sets (no MAML) on the ADIMA splits and compare per-language accuracy at k = 50, 100, 150, 200. If the plain classifier matches the reported 78–85% accuracies within noise, the claim that MAML enables the few-shot cross-lingual result is falsified, and the finding reduces to low-data classification on pre-trained features.

Watch

Extended reading notes

Core claim

The central claim is that pre-trained audio representations, when combined with MAML, yield a few-shot classifier that detects abusive speech across ten Indian languages with only tens to hundreds of labeled clips per language. Whisper's embeddings with $L_2$-norm feature normalization give the strongest results, with per-language accuracies between 78.98% and 85.22% in the 100-shot setting, and aggregate macro-F1 scores that exceed the ADIMA zero-shot baseline on four of the five languages for which that baseline reported numbers. The authors further claim that the $L_2$-norm consistently outperforms temporal-mean pooling for both feature extractors, and that a visual analysis of the features shows language-family clustering, with Dravidian languages forming distinct clusters and Indo-Aryan languages overlapping, which they interpret as evidence that phonetic similarity aids cross-lingual abuse detection.

Load-bearing premise

The reported results rest on an underspecified training protocol: the paper never states whether MAML is run with distinct support and query episodes or simply trained on a merged set of $k$ samples per language, and the claimed few-shot advantage depends on that distinction.

Editorial extensions

If this is right

With Whisper features and $L_2$-norm normalization, per-language accuracy sits between 78.98% and 85.22% at 100 shots, so roughly 100 labeled clips per language can support a usable abuse-detection pipeline.
Cross-lingual training appears to exploit acoustic family similarity: the feature study shows Dravidian languages clustering apart from Indo-Aryan languages, and Tamil–Malayalam plotting closest among the Dravidian group.
The $L_2$-norm normalization choice matters more than the choice between Whisper and Wav2Vec for final accuracy, since it improves both models' scores across most languages.
The method supplies macro-F1 numbers for all ten ADIMA languages, including seven that the original ADIMA zero-shot baseline did not report, and beats that baseline on four of the five languages it did report.

Reading between the lines

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

If the actual training procedure is a single model trained on the merged support set rather than episodic MAML with disjoint support and query sets, then MAML's contribution is not established; the result may be ordinary low-data classification on good features.
The language-family clustering in the feature visualization suggests a direct test: pretrain on Indo-Aryan languages and evaluate zero-shot on an unseen Dravidian language (or the reverse) to measure how much of the transfer is phonetic similarity rather than label-general features.
Because the pipeline uses frozen embeddings and a small network, it could be deployed as a lightweight moderation baseline for audio social platforms in under-resourced languages, but the unresolved training protocol would first need to be pinned down to make the few-shot claim reproducible.
The t-SNE overlap of Hindi-dialect languages (Bhojpuri, Haryanvi) with Hindi hints that the classifier may be leveraging dialect/acoustic similarity rather than abuse-specific cues for those languages; an adversarial evaluation on deliberately code-mixed or dialect-heavy audio would separate these explanations.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: the headline claim—MAML few-shot cross-lingual abuse detection—is not supported by the experimental write-up. What is actually new is the application of pre-trained audio features (Whisper, CLSRIL-23 Wav2Vec2) to the ADIMA dataset with a claimed few-shot setup, plus a comparison of L2-Norm vs Temporal Mean feature normalization. The per-language accuracy/F1 tables are useful, and the code is public. The t-SNE feature study is descriptive but not harmful.

The soft spots are load-bearing. Section 3.2 says "k samples per class" but then defines |S| = k × L and gives a k=2 example with one sample per class, making k total per language. Section 4.4 describes training an ANN with Adam for 150 epochs on the combined support set; there is no mention of task partitioning, inner-loop updates, or query sets. If that is what was implemented, the results are ordinary low-data classification on strong features, and the MAML label does no work. The "cross-lingual" claim is also weaker than stated: the support set is drawn from each language's own training split, so the model has seen in-language examples; testing is in-language. Comparing these numbers to ADIMA's zero-shot cross-lingual baseline in Table 4 is apples-to-oranges. There are also no variance estimates anywhere; a single run per setting is claimed.

None of this is fatal to the empirical direction. The idea of using pre-trained audio features for abuse detection in low-resource Indic languages is sensible, and Whisper with L2-Norm does seem to help. But as written, the paper does not establish that meta-learning contributes anything over a plain classifier. A careful revision that specifies the exact procedure, reports multiple seeds, and compares against a proper low-data baseline (including non-MAML training on the same support set) could turn this into a solid applied paper.

My recommendation: send it to peer review, but with a clear signal that the few-shot protocol is the crux. A referee should be asked to verify whether MAML is actually implemented and, if so, to insist on error bars and a real baseline comparison in the main text.

Referee Report

4 major / 6 minor

Summary. The paper proposes a few-shot audio abuse detection method for ten Indian languages using the ADIMA dataset. Audio features are extracted from Whisper and Wav2Vec (CLSRIL-23), normalized with either Temporal Mean or L2-Norm, and fed to a small neural network that, according to the paper, is trained within the Model-Agnostic Meta-Learning (MAML) framework. Experiments are reported at shot sizes 50, 100, 150, and 200, with accuracy and macro-F1 per language. The main claim is that Whisper with L2-Norm normalization achieves top accuracies of 78.98% to 85.22% in the 100-shot setting. A t-SNE feature visualization is also presented to support language-similarity observations.

Significance. If the claims hold, the paper would provide a practical recipe for low-resource audio moderation in a multilingual setting, and the comparison of feature extractors and normalization strategies would be a useful reference. The study is timely and addresses an underexplored modality. However, the current manuscript's central result is not firmly established because the training protocol is ambiguous and the reported numbers lack uncertainty estimates. The feature-visualization analysis is a strength in that it offers a qualitative, falsifiable claim about acoustic similarities across Dravidian and Indo-Aryan languages, but it does not compensate for the evaluation issues. The code release is a positive step for reproducibility, though the linked repository was not verified during review.

major comments (4)

[Section 3.2, Eq. (4)-(5) vs. Sections 4.3-4.4] The few-shot training protocol is not described consistently, and the role of MAML is unclear. Section 3.2 introduces 'k samples per class' but then Eq. (4) defines the support set with k total samples per language, and Eq. (5) sets |S| = k × L; the k=2 example states 'one abusive and one non-abusive sample per language,' which contradicts the standard k-shot-per-class terminology. More critically, Sections 4.3-4.4 describe training an ANN with Adam on the combined support set for 150 epochs but never specify the episodic MAML procedure: there is no mention of task sampling, inner-loop adaptation steps, task-specific learning rates in the inner loop, or query sets. If the implementation is ordinary supervised training on the merged support set, then 'MAML' and 'meta-learning' are labels that do no work, and the accuracy numbers simply reflect low-data classification on strong pre-trained features. This is load-bearing because the central claim of a few-shot meta-learning method rests on this description. The authors must provide an exact algorithmic description of the training procedure, including the inner- and outer-loop updates, or revise the claims accordingly.
[Section 5.1 and Tables 2-3] All reported accuracies and F1 scores come from a single experiment without variance estimates. Given that the support sets are randomly sampled (stratified by class) and the shot sizes are small, results such as Malayalam at 85.22% in the 100-shot Whisper/L2-Norm setting could be within run-to-run noise. The paper should report the mean and standard deviation (or confidence intervals) over multiple random seeds and multiple support-set draws, or at least state the number of repetitions if they were performed. Without this, the claims about the best normalization setting and the comparative ordering of languages are not statistically supported.
[Section 3.3 and Table 4] The evaluation protocol does not substantiate the 'cross-lingual' claim as it is used in the comparison. The model is trained on support sets from all ten languages and tested on test splits of the same ten languages, so each language contributes in-language training examples; this is multilingual few-shot classification, not cross-lingual generalization to an unseen language. Table 4 compares this to the ADIMA baselines of Gupta et al. (2022), which are zero-shot cross-lingual (train on a source language, test on a target language). This is an apples-to-oranges comparison. To support the cross-lingual claim, the paper should include a leave-one-language-out evaluation (training on some languages, testing on a held-out language) or explicitly reframe the contribution as multilingual in-language few-shot classification.
[Appendix A.2, Table 4] The aggregate macro-F1 baseline comparison is incomplete and nonstandard. Several ADIMA rows contain '-' with no explanation of whether those values are unavailable, not reported in the original paper, or not applicable. The table reports only aggregate macro-F1 for our best setting, while the language-specific tables (Tables 2-3) report both accuracy and F1; it is not clear how the aggregate values relate to the per-language numbers, and no statistical significance testing is provided. The authors should either complete the baseline rows with values from Gupta et al. (2022), or justify why those languages cannot be compared, and should clearly state the aggregation rule (e.g., macro-average over languages) when presenting the 'Aggregate Macro F1' column.

minor comments (6)

[Section 3.2, heading] The heading reads 'Meta Agnostic Meta Learning' and should be 'Model-Agnostic Meta-Learning'.
[Section 3.1, Eq. (2)] Equation (2) fixes the summation bound at 768, but Whisper features have dimension 1024 and Wav2Vec features have dimension 768. Please clarify whether Eq. (2) is intended for Wav2Vec only or should be written generically with the feature dimension as a variable.
[Figures 1 and 2] The heatmaps annotate cells with accuracy values but use a color scale that compresses the range 0.65-0.9, making differences such as 0.78 vs. 0.85 difficult to perceive; the appendix tables are helpful, but the figures would benefit from a narrower color range or a different annotation format.
[Section 5.1] The sentence 'with no top accuracy scores in the 200-shot scenario' is ambiguous; it should read 'no top accuracy scores occur in the 200-shot scenario' to match the surrounding discussion.
[Section 5.2, Figure 3] The t-SNE visualization is described without reporting hyperparameters such as perplexity, number of iterations, or learning rate, and without stating whether the projection was computed on the full dataset or a subsample. These details are needed for reproducibility.
[Section 5.1, Table 1] The dataset table shows class imbalance (e.g., Bhojpuri has 253 abusive vs. 506 non-abusive training samples). Since accuracy is the primary headline metric, the paper should explicitly discuss how imbalance affects the reported accuracies and why macro-F1, which is also reported, is not used for the headline claim.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reported accuracies are measured on held-out ADIMA test splits using externally pre-trained features and task labels; the one minor self-citation is motivational, not load-bearing.

full rationale

Walking the claimed derivation chain: the top accuracy scores (78.98-85.22% in the 100-shot setting) are empirical measurements on the ADIMA test splits, using features from externally pre-trained Whisper/Wav2Vec models and labels from the dataset. There is no fitted parameter that is later renamed a prediction, and no equation in the paper defines the outcome in terms of the input. The selection of L2-Norm normalization after observing its superior performance (Section 5.1, Figure 2b) is a model-selection step, not circular reasoning. The citation to Mozafari et al. (2022), which includes two of the present authors, appears only as motivational related work for meta-learning in cross-lingual abuse detection; the measured results do not depend on that citation, so it is not load-bearing. The MAML protocol inconsistency noted by the skeptic (Section 3.2 defines k-samples per class while Equation 5 gives |S| = k x L, and Sections 4.3-4.4 describe training on the combined support set without episodic task partitioning) is a correctness and reproducibility concern, not circularity: the reported accuracies are not constructed to equal the inputs. Similarly, comparing in-language few-shot results with ADIMA's zero-shot cross-lingual baseline in Table 4 is an evaluation-design issue rather than a circular derivation. Overall, the paper is self-contained against external benchmarks and contains no step where a prediction reduces by construction to its own input.

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

The paper introduces no new theoretical constructs. Its central claim depends on transfer from pre-trained models and on the correctness of the dataset. The main free parameters are standard hyperparameters, though epoch count was tuned on the task data. No invented entities.

free parameters (3)

meta and task learning rates = 0.001
Set to 0.001 for both learning rates in Section 4.4; no tuning procedure reported.
training epochs = 150
Chosen 'based on repeated testing' (Section 4.4), an ad hoc selection that may overfit the test set.
hidden layer sizes = 256 and 128
Architecture choice in Section 4.4, not justified by ablation.

assumptions (3)

domain assumption Pre-trained audio features from Whisper and Wav2Vec2 contain enough information to distinguish abusive from non-abusive speech.
The entire method relies on this; no fine-tuning or feature analysis is done beforehand (Section 3.1).
domain assumption MAML provides a benefit over simpler fine-tuning for this task.
The paper cites meta-learning successes in text and ASR but provides no comparison with a non-meta-learning baseline (Section 2.2, Section 4.3).
domain assumption ADIMA dataset labels and train/test splits are correct and representative.
Results are reported on these splits without auditing label quality (Section 4.1).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Cross-Lingual Audio Abuse Detection in Low-Resource Settings with Few-Shot Learning." pith.science (2026). https://pith.science/paper/XKID2JVN

@misc{pith2026241201408,
  author       = {Pith},
  title        = {Pith review of: Towards Cross-Lingual Audio Abuse Detection in Low-Resource Settings with Few-Shot Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XKID2JVN}},
  note         = {Machine review of arXiv:2412.01408}
}

read the original abstract

Online abusive content detection, particularly in low-resource settings and within the audio modality, remains underexplored. We investigate the potential of pre-trained audio representations for detecting abusive language in low-resource languages, in this case, in Indian languages using Few Shot Learning (FSL). Leveraging powerful representations from models such as Wav2Vec and Whisper, we explore cross-lingual abuse detection using the ADIMA dataset with FSL. Our approach integrates these representations within the Model-Agnostic Meta-Learning (MAML) framework to classify abusive language in 10 languages. We experiment with various shot sizes (50-200) evaluating the impact of limited data on performance. Additionally, a feature visualization study was conducted to better understand model behaviour. This study highlights the generalization ability of pre-trained models in low-resource scenarios and offers valuable insights into detecting abusive language in multilingual contexts.

Figures

Figures reproduced from arXiv: 2412.01408 by the authors.

**Figure 2.** L2-Norm: Few Shot Accuracies in 50, 100, 150 and 200 shot cases accuracy. Languages like Haryanvi, Punjabi, and Odia generally perform better than other languages across both models and normalization strategies. For Whisper with L2-Norm, Haryanvi has strong accuracy (84.7% at 100 shots), and similarly for Punjabi and Odia. Gujarati, Kannada, and Tamil show lower accuracy and F1 scores overall compared to others, re… view at source ↗

**Figure 3.** tSNE plot of L2-Norm Feature normalisation of Whisper Features extracted from the ADIMA dataset [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 34 canonical work pages

[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]

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]

Cleber Alc \^a ntara, Viviane Moreira, and Diego Feijo. 2020. https://aclanthology.org/2020.lrec-1.531 Offensive video detection: Dataset and baseline results . In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4309--4319, Marseille, France. European Language Resources Association

work page 2020
[4]

Md Rabiul Awal, Roy Ka-Wei Lee, Eshaan Tanwar, Tanmay Garg, and Tanmoy Chakraborty. 2024. https://doi.org/10.1109/TCSS.2023.3252401 Model-agnostic meta-learning for multilingual hate speech detection

arXiv 2024
[5]

Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations

work page 2020
[6]

Seamless Communication. 2023. https://arxiv.org/abs/2308.11596 Seamlessm4t: Massively multilingual & multimodal machine translation . Preprint, arXiv:2308.11596

arXiv 2023
[7]

Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdel rahman Mohamed, and Michael Auli. 2020. https://api.semanticscholar.org/CorpusID:220055837 Unsupervised cross-lingual representation learning for speech recognition

work page 2020
[8]

Marta Costa-juss \`a , Mariano Meglioli, Pierre Andrews, David Dale, Prangthip Hansanti, Elahe Kalbassi, Alexandre Mourachko, Christophe Ropers, and Carleigh Wood. 2024. https://aclanthology.org/2024.findings-acl.340 M u T ox: Universal MU ltilingual audio-based TOX icity dataset and zero-shot detector . In Findings of the Association for Computational Li...

work page 2024

Show all 56 references

[9]

Dar and Dolly Nagrath

S. Dar and Dolly Nagrath. 2022. https://doi.org/10.22161/ijels.74.8 The impact that social media has had on today's generation of indian youth: An analytical study

2022 doi
[10]

Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. https://arxiv.org/abs/1703.03400 Model-agnostic meta-learning for fast adaptation of deep networks . Preprint, arXiv:1703.03400

2017 arXiv
[11]

S. G. 2022. https://doi.org/10.34256/irjt22s533 Family relations in the moral values expressed by dravidian literature

2022 doi
[12]

Kumaraguru

Niloy Ganguly and P. Kumaraguru. 2019. https://doi.org/10.1145/3345671 The positive and negative effects of social media in india

2019 doi
[13]

Zhiwei Gao, Shuntaro Yada, Shoko Wakamiya, and Eiji Aramaki. 2020. https://doi.org/10.18653/v1/2020.coling-main.175 Offensive language detection on video live streaming chat . In Proceedings of the 28th International Conference on Computational Linguistics, pages 1936--1940, B...

2020 doi
[14]

G. V. Garje, A. Bansode, Suyog Gandhi, and Adita Kulkarni. 2016. https://doi.org/10.5120/IJCA2016908837 Marathi to english sentence translator for simple assertive and interrogative sentences

2016 doi
[15]

Sreyan Ghosh, Samden Lepcha, Sahni Sakshi, Rajiv Ratn Shah, and Srinivasan Umesh. 2021. Detoxy: A large-scale multimodal dataset for toxicity classification in spoken utterances

2021
[16]

Sascha Grollmisch, Estefan \' a Cano, Christian Kehling, and Michael Taenzer. 2021. Analyzing the potential of pre-trained embeddings for audio classification tasks. In 2020 28th European Signal Processing Conference (EUSIPCO), pages 790--794. IEEE

2021
[17]

Jiatao Gu, Yong Wang, Yun Chen, Kyunghyun Cho, and V. Li. 2018. https://doi.org/10.18653/v1/D18-1398 Meta-learning for low-resource neural machine translation

2018 doi
[18]

Anirudh Gupta, Harveen Singh Chadha, Priyanshi Shah, Neeraj Chhimwal, Ankur Dhuriya, Rishabh Gaur, and Vivek Raghavan. 2021. Clsril-23: Cross lingual speech representations for indic languages

2021
[19]

Gupta, R

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

2022
[20]

Wenxin Hou, Yidong Wang, Shengzhou Gao, and Takahiro Shinozaki. 2021 a . https://api.semanticscholar.org/CorpusID:235780558 Meta-adapter: Efficient cross-lingual adaptation with meta-learning

2021
[21]

Wenxin Hou, Hanlin Zhu, Yidong Wang, Jindong Wang, Tao Qin, Renjun Xu, and Takahiro Shinozaki. 2021 b . https://api.semanticscholar.org/CorpusID:235187234 Exploiting adapters for cross-lingual low-resource speech recognition

2021
[22]

Jui-Yang Hsu, Yuan-Jui Chen, and Hung yi Lee. 2019. https://api.semanticscholar.org/CorpusID:204904570 Meta learning for end-to-end low-resource speech recognition

2019
[23]

Taehun Huh and Youngjoong Ko. 2022. https://doi.org/10.1145/3477495.3531908 Lightweight meta-learning for low-resource abstractive summarization

2022
[24]

Danesh Jain and George Cardona. 2007. https://doi.org/10.4324/9780203945315 The indo-aryan languages

2007 doi
[25]

Jaswanth, N

M. Jaswanth, N. V. L. Narayana, Sreedharreddy Rahul, Susmitha Vekkot, and Sreedharreddy Rahul. 2022. https://doi.org/10.1109/CENTCON56610.2022.10051465 A comparative study of feature modelling methods for telugu language identification

2022
[26]

Margam, Pooja Aggarwal, R

Simran Khanuja, Diksha Bansal, Sarvesh Mehtani, Savya Khosla, Atreyee Dey, Balaji Gopalan, D. Margam, Pooja Aggarwal, R. Nagipogu, Shachi Dave, Shruti Gupta, Subhash Chandra Bose Gali, Vishnu Subramanian, and P. Talukdar. 2021. Muril: Multilingual representations for indian languages

2021
[27]

Kingma and Jimmy Ba

Diederik P. Kingma and Jimmy Ba. 2017. https://arxiv.org/abs/1412.6980 Adam: A method for stochastic optimization . Preprint, arXiv:1412.6980

2017 arXiv
[28]

Plumbley

Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D. Plumbley. 2020. https://arxiv.org/abs/1912.10211 Panns: Large-scale pretrained audio neural networks for audio pattern recognition . Preprint, arXiv:1912.10211

2020 arXiv
[29]

A. Menon. 1990. https://doi.org/10.1017/S0041977X00021285 Some observations on the sub-group tamil-malayalam: differential realizations of the cluster *nt

1990 doi
[30]

Don P Mitchell. 1996. Consequences of stratified sampling in graphics. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 277--280

1996
[31]

Springer

Marzieh Mozafari, Reza Farahbakhsh, and Noel Crespi. 2019. https://arxiv.org/abs/1910.12574 A bert-based transfer learning approach for hate speech detection in online social media . In "Springer"

2019 arXiv
[32]

Marzieh Mozafari, Reza Farahbakhsh, and Noel Crespi. 2022. https://doi.org/10.1109/ACCESS.2022.3147588 Cross-lingual few-shot hate speech and offensive language detection using meta learning

2022
[33]

Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Noboru Harada, and Kunio Kashino. 2022. Byol for audio: Exploring pre-trained general-purpose audio representations

2022
[34]

Shriphani Palakodety and Ashiqur KhudaBukhsh. 2020. https://doi.org/10.18653/v1/2020.wnut-1.24 Annotation efficient language identification from weak labels . In Proceedings of the Sixth Workshop on Noisy User-generated Text (W-NUT 2020), pages 181--192, Online. Association fo...

2020 doi
[35]

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

2019
[36]

Orchid Chetia Phukan, Yashasvi Chaurasia, Arun Balaji Buduru, and Rajesh Sharma. 2024. Collab: A collaborative approach for multilingual abuse detection

2024
[37]

Flor Miriam Plaza-Del-Arco, M Dolores Molina-Gonz \'a lez, L Alfonso Ure \ n a-L \'o pez, and Mar \' a Teresa Mart \' n-Valdivia. 2021. A multi-task learning approach to hate speech detection leveraging sentiment analysis

2021
[38]

Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. https://arxiv.org/abs/2212.04356 Robust speech recognition via large-scale weak supervision . Preprint, arXiv:2212.04356

2022 arXiv
[39]

Shantanu Kumar Rahut, Riffat Sharmin, and Ridma Tabassum. 2020. https://doi.org/10.1109/ETCCE51779.2020.9350919 Bengali abusive speech classification: A transfer learning approach using vgg-16 . In 2020 Emerging Technology in Computing, Communication and Electronics (ETCCE), p...

2020
[40]

Aneri Rana and Sonali Jha. 2022. https://arxiv.org/abs/2202.06218 Emotion based hate speech detection using multimodal learning . Preprint, arXiv:2202.06218

2022 arXiv
[41]

Tharindu Ranasinghe and Marcos Zampieri. 2021. https://doi.org/10.1145/3457610 Multilingual offensive language identification for low-resource languages

2021 doi
[42]

Aaqib Saeed, David Grangier, and Neil Zeghidour. 2021. Contrastive learning of general-purpose audio representations. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3875--3879. IEEE

2021
[43]

Sengupta

S. Sengupta. 2018. https://doi.org/10.4103/ijo.IJO_587_18 Vision assessment in regional indian languages

2018 doi
[44]

Sharon, H

R. Sharon, H. Shah, D. Mukherjee, and V. Gupta. 2022. Multilingual and multimodal abuse detection

2022
[45]

Rini Sharon and Debdoot Mukherjee. 2024. https://doi.org/10.1109/ICASSP48485.2024.10447259 Study of abuse detection in continuous speech for indian languages . In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 11756--11760

2024
[47]

Satwinder Singh, Ruili Wang, and Feng Hou. 2022. https://doi.org/10.1109/ICASSP43922.2022.9746899 Improved meta learning for low resource speech recognition

2022
[48]

Sinha, Aruna Jain, and S

S. Sinha, Aruna Jain, and S. Agrawal. 2014. https://doi.org/10.1007/978-3-319-04960-1_14 Speech processing for hindi dialect recognition

2014 doi
[49]

Jake Snell, Kevin Swersky, and Richard S. Zemel. 2017. https://arxiv.org/abs/1703.05175 Prototypical networks for few-shot learning . Preprint, arXiv:1703.05175

2017 arXiv
[50]

Spiesberger, Andreas Triantafyllopoulos, Iosif Tsangko, and Björn W

Anika A. Spiesberger, Andreas Triantafyllopoulos, Iosif Tsangko, and Björn W. Schuller. 2023. https://doi.org/10.21437/Interspeech.2023-789 Abusive Speech Detection in Indic Languages Using Acoustic Features . In Proc. INTERSPEECH 2023, pages 2683--2687

2023 doi
[51]

Srivastava, S

K. Srivastava, S. Chaudhury, J. Prakash, and Sana Dhamija. 2019. https://doi.org/10.4103/ipj.ipj_154_20 Social media and mental health challenges

2019 doi
[52]

Srivatsa and Annarao Kulkarni

B. Srivatsa and Annarao Kulkarni. 2017. https://doi.org/10.1109/ICACCAF.2017.8344737 The curious case of kannada ‘maadu’

2017
[53]

Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, and Hugo Larochelle. 2020. https://openreview.net/forum?id=rkgAGAVKPr Meta-dataset: A dataset of datasets for learning ...

2020
[54]

Niels van der Heijden, Helen Yannakoudakis, Pushkar Mishra, and Ekaterina Shutova. 2021. https://doi.org/10.18653/v1/2021.eacl-main.168 Multilingual and cross-lingual document classification: A meta-learning approach . In Proceedings of the 16th Conference of the European Chap...

2021 doi
[55]

Yuanyuan Wang, Yu Gu, Yifei Yin, Yingping Han, He Zhang, Shuang Wang, Chenyu Li, and Dou Quan. 2023. Multimodal transformer augmented fusion for speech emotion recognition

2023
[56]

Mengzhou Xia, Guoqing Zheng, Subhabrata Mukherjee, Milad Shokouhi, Graham Neubig, and Ahmed Hassan Awadallah. 2021. https://doi.org/10.18653/v1/2021.naacl-main.42 M eta XL : Meta representation transformation for low-resource cross-lingual learning . In Proceedings of the 2021...

2021 doi
[57]

Zhanyuan Yang, Jinghua Wang, and Yingying Zhu. 2022. https://arxiv.org/abs/2209.08224 Few-shot classification with contrastive learning . Preprint, arXiv:2209.08224

2022 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[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]

Cleber Alc \^a ntara, Viviane Moreira, and Diego Feijo. 2020. https://aclanthology.org/2020.lrec-1.531 Offensive video detection: Dataset and baseline results . In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4309--4319, Marseille, France. European Language Resources Association

work page 2020

[4] [4]

Md Rabiul Awal, Roy Ka-Wei Lee, Eshaan Tanwar, Tanmay Garg, and Tanmoy Chakraborty. 2024. https://doi.org/10.1109/TCSS.2023.3252401 Model-agnostic meta-learning for multilingual hate speech detection

arXiv 2024

[5] [5]

Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations

work page 2020

[6] [6]

Seamless Communication. 2023. https://arxiv.org/abs/2308.11596 Seamlessm4t: Massively multilingual & multimodal machine translation . Preprint, arXiv:2308.11596

arXiv 2023

[7] [7]

Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdel rahman Mohamed, and Michael Auli. 2020. https://api.semanticscholar.org/CorpusID:220055837 Unsupervised cross-lingual representation learning for speech recognition

work page 2020

[8] [8]

Marta Costa-juss \`a , Mariano Meglioli, Pierre Andrews, David Dale, Prangthip Hansanti, Elahe Kalbassi, Alexandre Mourachko, Christophe Ropers, and Carleigh Wood. 2024. https://aclanthology.org/2024.findings-acl.340 M u T ox: Universal MU ltilingual audio-based TOX icity dataset and zero-shot detector . In Findings of the Association for Computational Li...

work page 2024