Elamur Research⌕
☰
home / papers / 2026 / “OK Aura, Be Fair with Me”: Demographics-Agnostic Training for Bias Mitigation in Wake-up Word Detection

“OK Aura, Be Fair with Me”: Demographics-Agnostic Training for Bias Mitigation in Wake-up Word Detection

Paper · 2026 · Fernando López, Paula Delgado-Santos, Pablo Gómez et al.

Summary

Wake-up Word (WuW) detectors are the always-on entry point to voice assistants. Their decisions rest on very short audio snippets, so they can treat speakers of different sex, age, and accent unequally, forcing some users to repeat the trigger phrase more often. Most bias-mitigation methods need demographic labels, which are often missing, incomplete, or privacy-sensitive.

The authors therefore train WuW models without any demographic information and use the labels only to audit the results afterwards. Working with the Spanish OK Aura dataset, they first locate underrepresented groups and measure bias with Predictive Disparity (PD) and Disparate Impact (DI). They then test two label-free strategies:

  • Spectral data augmentations, meant to disrupt acoustic cues that correlate with demographics, such as F0, formants, and prosody.
  • Knowledge distillation, where a large frozen self-supervised teacher passes its knowledge to a compact student model.

Both strategies reduce group disparities noticeably while keeping overall detection quality. The best method cuts PD by roughly 40–84% depending on the attribute. This suggests fairer on-device WuW systems are achievable without collecting sensitive personal data.

Key points

  • The dataset is OK Aura (Okey Aura Wake-up Word Dataset, Zenodo), a real-world Spanish WuW corpus. Sex, age, and accent labels are used for evaluation only, never for training.
  • The augmentations tested are FreqMixStyle, FilterAugment (linear variant), SpecAugment-style frequency masking, and Device Impulse Responses (DIR).
  • The teacher model, w2v-BERT2-kws, is built on the frozen w2v-BERT 2.0 encoder, a 24-layer Conformer pre-trained on 4.5M hours of audio. Only its classification head is trained:
    • a learnable weighted sum of the layer outputs,
    • Multi-Head Factorized Attention (MHFA),
    • attentive pooling,
    • a linear classifier.
  • The teacher is used only for distillation and is not intended for on-device inference.
  • Relative to the baseline, the best technique lowers PD by 39.94% (sex), 83.65% (age), and 40.48% (accent).

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

🎓Universidad Autónoma de MadridMadrid · Spain7 works · 21 citations

Also on the byline: Telefónica · Telefónica Innovación Digital.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksSafety Research · 543 works

Cite

Cite as Fernando López, Paula Delgado-Santos, Pablo Gómez, David Solans and Jordi Luque. 2026. “OK Aura, Be Fair with Me”: Demographics-Agnostic Training for Bias Mitigation in Wake-up Word Detection. Proceedings of Speech Language Models in Low-Resource Settings: Performance, Evaluation, and Bias Analysis (SPEAKABLE) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3sgj2rai4uv8

@inproceedings{lopez2026ok,
  title = {{“OK Aura, Be Fair with Me”: Demographics-Agnostic Training for Bias Mitigation in Wake-up Word Detection}},
  author = {Fernando López, Paula Delgado-Santos, Pablo Gómez, David Solans and Jordi Luque},
  booktitle = {Proceedings of Speech Language Models in Low-Resource Settings: Performance, Evaluation, and Bias Analysis (SPEAKABLE) @ LREC 2026},
  year = {2026},
  doi = {10.63317/3sgj2rai4uv8},
  url = {https://doi.org/10.63317/3sgj2rai4uv8}
}

References

52 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Giuseppe Attanasio, Beatrice Savoldi, Dennis Fucci, and Dirk Hovy. 2024. 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, U… find
  2. Marcio Fuckner, Sophie Horsman, Pascal Wiggers, and Iskaj Janssen. 2023. Uncovering bias in asr systems: Evaluating wav2vec2 and whisper for dutch speakers. In IEEE International Conference on Speech Technology and Human-Computer Dialogue (SpeD), pages 146–151. find
  3. Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. Word embeddings quantify 100 years of gender and ethnic stereotypes. Proceedings of the National Academy of Sciences, 115(16):E3635–E3644. find
  4. Andrew Bailey and Mark D Plumbley. 2021. Gender bias in depression detection using audio features. In IEEE 29th European Signal Processing Conference (EUSIPCO), pages 596–600. find
  5. Jiangyu Han, Federico Landini, Johan Rohdin, Anna Silnova, Mireia Diez, and Lukáš Burget. 2025. Leveraging self-supervised learning for speaker diarization. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. find
  6. Solon Barocas and Andrew D. Selbst. 2016. Big data’s disparate impact. California Law Review, 104(3):671–732. find
  7. Loïc Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Mark Duppenthaler, Paul-Ambroise Duquenne, Brian Ellis, Hady Elsahar, Justin Haaheim, et al. 2023. Seamless: Multilingual expressive and streaming speech translation. arXiv preprint arXiv:2312.05187.
  8. James D Harnsberger, Rahul Shrivastav, William S Brown Jr, Howard Rothman, and Harry Hollien. 2008. Speaking rate and fundamental frequency as speech cues to perceived age. Journal of voice, 22(1):58–69. find
  9. Junyi Chai, Zhihao Wang, Jiaxin Chen, Hao He, Dawn Song, and Xia Li. 2022. Fairness without demographics through knowledge distillation. In Advances in Neural Information Processing Systems (NeurIPS), volume 35. find
  10. Camille Harris, Chijioke Mgbahurike, Neha Kumar, and Diyi Yang. 2024. Modeling gender and dialect bias in automatic speech recognition. In Findings of the Association for Computational Linguistics: EMNLP, pages 15166–15184.
  11. Zhengyang Chen, Sanyuan Chen, Yu Wu, Yao Qian, Chengyi Wang, Shujie Liu, Yanmin Qian, and Michael Zeng. 2022. Large-scale self-supervised speech representation learning for automatic speaker verification. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 614… find
  12. Wiebke Hutiri, Aaron Yi Ding, Fahim Kawsar, and Akhil Mathur. 2023. Tiny, always-on, and fragile: Bias propagation through design choices in on-device machine learning workflows. ACM Transactions on Software Engineering and Methodology, 32(6):1–37. find
  13. Anna Seo Gyeong Choi and Hoon Choi. 2025. Fairness of automatic speech recognition: Looking through a philosophical lens. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(1):605–614. find
  14. Veton Këpuska and Jason Breitfeller. 2006. Wake-up-word speech recognition application for first responder communication enhancement. In Sensors, and Command, Control, Communications, Yen Meng, Yi-Hui Chou, Andy T. Liu, and Hung-yi Lee. 2022. Don’t speak too fast: The impact of data bias on self-sup… find
  15. Byeonggeun Kim, Seunghan Yang, Jangho Kim, Hyunsin Park, Juntae Lee, and Simyung Chang. 2022. Domain generalization with relaxed instance frequency-wise normalization for multi-device acoustic scene classification. arXiv preprint arXiv:2206.12513. — [arXiv:2206.12513](https://arxiv.org/abs/2206.12513) find
  16. Abdelrahman Mohamed, Hung-yi Lee, Lasse Borgholt, Jakob D. Havtorn, Joakim Edin, Christian Igel, Katrin Kirchhoff, Shang-Wen Li, Karen Livescu, Lars Maaløe, Tara N. Sainath, and Shinji Watanabe. 2022. Self-supervised speech representation learning: A review. IEEE Journal of Selected Topics in Signal… find
  17. Gwantae Kim, David K Han, and Hanseok Ko. 2021. Specmix: A mixed sample data augmentation method for training withtime-frequency domain features. arXiv preprint arXiv:2108.03020. — [arXiv:2108.03020](https://arxiv.org/abs/2108.03020) find
  18. Jacob Mørk, Holger Severin Bovbjerg, Gergely Kiss, and Zheng-Hua Tan. 2024. Noise-robust keyword spotting through self-supervised pre-training. arXiv preprint arXiv:2403.18560. — [arXiv:2403.18560](https://arxiv.org/abs/2403.18560) find
  19. Alkis Koudounas, Flavio Giobergia, Eliana Pastor, and Elena Baralis. 2024. A contrastive learning approach to mitigate bias in speech models. arXiv preprint arXiv:2406.14686. — [arXiv:2406.14686](https://arxiv.org/abs/2406.14686) find
  20. Tobias Morocutti, Florian Schmid, Khaled Koutini, and Gerhard Widmer. 2023. Device-robust acoustic scene classification via impulse response augmentation. In IEEE 31st European Signal Processing Conference (EUSIPCO), pages 176–180. find
  21. Veton Këpuska. 2011. Wake-up-word speech recognition. In Ivo Ipsic, editor, Speech Technologies, chapter 12. IntechOpen, London. find
  22. Beltrán Labrador, Pai Zhu, Guanlong Zhao, Angelo Scorza Scarpati, Quan Wang, Alicia Lozano-Diez, and Ignacio Lopez-Moreno. 2025. Personalizing keyword spotting with speaker information. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. find
  23. Dena Mujtaba, Nihar Mahapatra, Megan Arney, J Yaruss, Hope Gerlach-Houck, Caryn Herring, and Jia Bin. 2024. Lost in transcription: Identifying and quantifying the accuracy biases of automatic speech recognition systems against disfluent speech. In Proceedings of Conference of the North American Chap… find
  24. Yi-Cheng Lin, Tzu-Quan Lin, Hsi-Che Lin, Andy T Liu, and Hung-yi Lee. 2024. On the social bias of speech self-supervised models. In Proceedings of INTERSPEECH, pages 4638–4642. find
  25. Fernando López and Jordi Luque. 2024. Albayzin evaluation 2024: Wake-up word detection challenge. find
  26. Hyeonuk Nam, Seong-Hu Kim, and Yong-Hwa Park. 2022. Filteraugment: An acoustic environmental data augmentation method. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). find
  27. Iván López-Espejo, Zheng-Hua Tan, John HL Hansen, and Jesper Jensen. 2021. Deep spoken keyword spotting: An overview. IEEE Access, 10:4169–4199. find
  28. Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. 2019. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779.
  29. Fernando López and Jordi Luque. 2022. Iterative pseudo-forced alignment by acoustic CTC loss for self-supervised ASR domain adaptation. In Proceedings of IberSPEECH, pages 46–50. find
  30. Ankita Pasad, Bowen Shi, and Karen Livescu. 2023. Comparative layer-wise analysis of self-supervised speech models. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. find
  31. Fernando López, Jordi Luque, Carlos Segura, and Pablo Gómez. 2023. Robust wake-up word detection by two-stage multi-resolution ensembles. arXiv preprint arXiv:2310.11379. — [arXiv:2310.11379](https://arxiv.org/abs/2310.11379) find
  32. Eliana Pastor, Luca De Alfaro, and Elena Baralis. 2021. Looking for trouble: Analyzing classifier behavior via pattern divergence. In Proceedings of the International Conference on Management of Data, pages 1400–1412. find
  33. Joshua L Martin and Kelly Elizabeth Wright. 2023. Bias in automatic speech recognition: The case of african american language. Applied Linguistics, 44(4):613–630. find
  34. Junyi Peng, Ladislav Mošner, Lin Zhang, Oldřich Plchot, Themos Stafylakis, Lukáš Burget, and Jan Černockỳ. 2025. Ca-mhfa: A context-aware multi-head factorized attentive pooling for ssl-based speaker verification. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),… find
  35. Maryam Zolnoori, Sasha Vergez, Zidu Xu, Elyas Esmaeili, Ali Zolnour, Krystal Anne Briggs, Jihye Kim Scroggins, Seyed Farid Hosseini Ebrahimabad, James M Noble, Maxim Topaz, et al. 2024. Decoding disparities: evaluating automatic speech recognition system performance in transcribing black and white p… find
  36. Marina Piat, Dominique Fohr, and Irina Illina. 2008. Foreign accent identification based on prosodic parameters. In INTERSPEECH, pages 759–762. 9. find
  37. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning (ICML), volume 202, pages 28492–28518. PMLR.
  38. Language Resource References Guillermo Cámbara, Jordi Luque, David Bonet, Fernando López, Mireia Farrús, Pablo Gómez and Carlos Segura. 2024. Okey Aura Wake-up Word Dataset. Zenodo, 1.1.0. find
  39. Daniel Roncel Díaz, Federico Costa, and Javier Hernando. 2024. On the use of audio to improve dialogue policies. In IberSPEECH, pages 151– 155. find
  40. Maliha Jahan, Yinglun Sun, Priyam Mazumdar, Zsuzsanna Fagyal, Thomas Thebaud, Jesus Villalba, Mark Hasegawa-Johnson, Najim Dehak, and Laureano Moro Velazquez. 2025. Faist: A benchmark dataset for fairness in speech technology. In Proceedings of Interspeech, pages 1343–1347. find
  41. Harvineet Singh, Fan Xia, Mi-Ok Kim, Romain Pir-racchio, Rumi Chunara, and Jean Feng. 2023. A brief tutorial on sample size calculations for fairness audits. arXiv preprint arXiv:2312.04745. — [arXiv:2312.04745](https://arxiv.org/abs/2312.04745) find
  42. Joachim Thiemann, Nobutaka Ito and Emmanuel Vincent. 2013. DEMAND: Diverse Environments Multi-Channel Acoustic Noise Database. The Journal of the Acoustical Society of America. find
  43. Isaac Slaughter, Craig Greenberg, Reva Schwartz, and Aylin Caliskan. 2023. Pre-trained speech processing models contain human-like biases that propagate to speech emotion recognition. pages 8967–8989. find
  44. Juan Carlos Franco Hernández, Tim Brookes, Enzo De Sena. 2021. Multi-Angle, Multi-Distance Microphone Impulse Response Dataset. Zenodo, 1.0.0. find
  45. Loes Vandenberghe et al. 2023. Exploring data augmentation in bias mitigation against non-native-accented speech. arXiv preprint arXiv:2312.15499. — [arXiv:2312.15499](https://arxiv.org/abs/2312.15499) find
  46. Mozilla Foundation. 2021. Common Voice Corpus (Spanish), version 7.1. Mozilla Common Voice. find
  47. Houri K Vorperian, Raymond D Kent, Yen Lee, and Daniel M Bolt. 2019. Corner vowels in males and females ages 4 to 20 years: Fundamental and f1–f4 formant frequencies. The Journal of the Acoustical Society of America, 146(5):3255– 3274. find
  48. OpenSLR. 2016. Room Impulse Response and Noise Database (SLR28). OpenSLR. find
  49. Imdat Solak. 2019. The M-AILABS Speech Dataset. M-AILABS. find
  50. Irina-Elena Veliche, Zhuangqun Huang, Vineeth Ayyat Kochaniyan, Fuchun Peng, Ozlem Kalinli, and Michael L Seltzer. 2024. Towards measuring fairness in speech recognition: Fair-speech dataset. arXiv preprint arXiv:2408.12734. — [arXiv:2408.12734](https://arxiv.org/abs/2408.12734) find
  51. Xin Wang, Héctor Delgado, Hemlata Tak, Jee-weon Jung, Hye-jin Shim, Massimiliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidullah, Tomi Kinnunen, et al. 2024. Asvspoof 5: Crowdsourced speech data, deepfakes, and adversarial attacks at scale. arXiv preprint arXiv:2408.08739. — [arXiv:2408.08739](https://arxiv.org/abs/2408.08739) find
  52. Mingdong Yu, Xiaofeng Jin, Bangxian Wan, and Guirong Wang. 2023. A few-shot speech keyword spotting method based on self-supervise learning. In 16th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), pages 1– 5. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/lopez-2026-ok-aura-be-fair-me/