Elamur Research⌕
☰
home / papers / 2026 / Speak No Evil, Just Prompt: Low-resource Multilingual Toxic Speech Detection with Audio Language Model

Speak No Evil, Just Prompt: Low-resource Multilingual Toxic Speech Detection with Audio Language Model

Paper · 2026 · Mingzi Zuo, Lei Zhang, Hailiang Sun et al.

Summary

Toxic speech detection has received far less attention than its text counterpart. Most existing systems run speech through ASR and then a text classifier. This setup breaks down for low-resource languages and accents, where transcription errors are frequent, and it loses acoustic signals such as prosody. To close the data gap, the authors build PolySpeechTox, a speech corpus labeled toxic/safe that spans 53 languages and accent varieties. It mixes real user voice comments with filtered TTS speech, and each label requires unanimous agreement among three annotators. They also introduce SoftPrompt-TSD, which detects toxicity end-to-end without ASR. It keeps an audio language model (MiMo-Audio-7B) frozen and trains only soft prompts of three kinds: a task prompt, a prompt shared across languages, and a per-language/accent residual prompt. The method reaches a micro ROC-AUC of 98.07% and outperforms baselines and other PEFT methods. It also transfers better to unseen low-resource languages, held-out accents and a different corpus (MuTox). The work offers a resource and a lightweight approach for multilingual speech moderation beyond high-resource languages.

Key points

  • PolySpeechTox: 53 languages and accent varieties, including low-resource languages such as Swahili, Sinhala and Zulu, plus accent variants of English, Arabic and Chinese.
    • Synthetic clips are kept only if three listeners all judge them to be human speech.
    • Annotations are checked by a separate quality-control group using agreement rate and Cohen's κ.
  • Comparison with other PEFT baselines: SoftPrompt-TSD scores 98.07 micro-AUC, versus 94.04 for LoRA and 92.92 for prefix tuning on the same frozen backbone.
  • Ablation:
    • The frozen ALM with only a fixed text instruction scores 55.49 AUC, barely above chance.
    • Removing the language-specific residual prompt causes the largest drop, to 95.58.
  • Cross-accent transfer (AUC): SoftPrompt-TSD beats the strongest baseline for each language: Chinese 81.00 vs 75.36, English 74.37 vs 69.06, Arabic 82.74 vs 76.29.
  • Zero-shot and out-of-domain tests: In zero-shot transfer from high- or medium-resource languages to low-resource ones, and in cross-corpus tests with MuTox, the method degrades less than the baselines.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Tianjin UniversityTianjin · China52 works · 1,128 citations🎓Shenzhen UniversityShenzhen · China22 works · 95 citations🎓Guangzhou UniversityGuangzhou · China6 works · 7 citations🎓Qilu University of TechnologyJinan · China7 works · 3 citations🤝National Computer Network Emergency Response Technical Team/Coordination Center of ChinarBeijing · China2 works

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksSignal Processing · 493 works

🧭 Topics 4Research subjects and languages the work is about.Multilingual & cross-lingual NLP · 882 worksBias, fairness & safety · 587 worksSpeech & spoken language · 519 worksLow-resource languages · 287 works

Cite

Cite as Mingzi Zuo, Lei Zhang, Hailiang Sun, Shengzhi Huo, Changyu Dong, Xin Wang, Bo Wang and Hao Liu. 2026. Speak No Evil, Just Prompt: Low-resource Multilingual Toxic Speech Detection with Audio Language Model. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.439

@inproceedings{zuo2026speak,
  title = {{Speak No Evil, Just Prompt: Low-resource Multilingual Toxic Speech Detection with Audio Language Model}},
  author = {Mingzi Zuo, Lei Zhang, Hailiang Sun, Shengzhi Huo, Changyu Dong, Xin Wang, Bo Wang and Hao Liu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.439},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.439}
}

References

10 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. Common voice: A massively-multilingual speech corpus. In Proceedings of the twelfth language resources and evaluation conference, pages 421…
  2. Kirtilekha Bhesra and Akshay Agarwal. 2024. A multimodal framework to counter hate speeches. In International Conference on Pattern Recognition, pages 197–207. Springer. find
  3. Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shiliang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou. 2023. Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models. arXiv preprint arXiv:2311.07919.
  4. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tai, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, and 16 others. 2024.… find
  5. Marta Costa-jussà, Mariano Meglioli, Pierre Andrews, David Dale, Prangthip Hansanti, Elahe Kalbassi, Alexandre Mourachko, Christophe Ropers, and Carleigh Wood. 2024. Mutox: Universal multilingual audio-based toxicity dataset and zero-shot detector. In Findings of the Association for Computational Li… find
  6. Huy Ba Do, Vy Le-Phuong Huynh, and Luan Thanh Nguyen. 2025. Vitosa: Audio-based toxic spans detection on vietnamese speech utterances. In Interspeech 2025, pages 4013–4017. find
  7. Sreyan Ghosh, Samden Lepcha, S Sakshi, Rajiv Ratn Shah, and Srinivasan Umesh. 2022. DeToxy: A Large-Scale Multimodal Dataset for Toxicity Classification in Spoken Utterances. In Interspeech 2022, pages 5185–5189. find
  8. Vikram Gupta, Rini Sharon, Ramit Sawhney, and Debdoot 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. find
  9. Janosch Haber, Bertie Vidgen, Matthew Chapman, Vibhor Agarwal, Roy Ka-Wei Lee, Yong Keong Yap, and Paul Röttger. 2023. Improving the detection of multilingual online attacks with rich social media data from Singapore. In Proceedings of the 61st Annual Meeting of the Association for Computational Lin… find
  10. Hareem Kibriya, Ayesha Siddiqa, Wazir Zada Khan, and Muhammad Khurram Khan. 2024. Towards safer online communities: Deep learning and explainable AI for hate speech detection and classification. Computers and Electrical Engineering, 116:109153. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zuo-2026-speak-no-evil-just-prompt/