--- title: "Speak Your Mind — paper by Satish et al. (2026)" description: "The paper asks whether speech foundation models carry voice-based social bias when continuing a spoken prompt. 3 references." authors: ["Shree Harsha Bokkahalli Satish", "Harm Lameris", "Olivier Perrotin", "Gustav Eje Henter", "Eva Szekely"] year: 2026 venue: "Proceedings of the Second Workshop of Identity Aware AI" source: "https://aclanthology.org/2026.iaai-1.2/" doi: "10.63317/3h9cs6yh6gvq" licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)" licence_url: "https://creativecommons.org/licenses/by-nc/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/satish-2026-speak-your-mind-speech-continuation/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/satish-2026-speak-your-mind-speech-continuation/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/43/) / Speak Your Mind: The Speech Continuation Task as a Probe of Voice-Based Model Bias # Speak Your Mind: The Speech Continuation Task as a Probe of Voice-Based Model Bias **Authors:** Shree Harsha Bokkahalli Satish, [Harm Lameris](https://md.research.elamur.ai/free/en/authors/harm-lameris/), Olivier Perrotin, Gustav Eje Henter and [Eva Szekely](https://md.research.elamur.ai/free/en/authors/eva-szekely/) · **Year:** 2026 · **Venue:** [Proceedings of the Second Workshop of Identity Aware AI](https://md.research.elamur.ai/free/en/venues/iaai/) ([Workshop of Identity Aware AI 2026](https://md.research.elamur.ai/free/en/venues/iaai/2026/)) · **Publisher:** [European Language Resources Association (ELRA)](https://md.research.elamur.ai/free/en/publishers/european-language-resources-association-elra/) · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.iaai-1.2/) · [PDF](https://aclanthology.org/2026.iaai-1.2.pdf) · [DOI 10.63317/3h9cs6yh6gvq](https://doi.org/10.63317/3h9cs6yh6gvq) · [OpenAlex](https://openalex.org/W4416432130) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary The paper asks whether speech foundation models carry voice-based social bias when continuing a spoken prompt. Speech continuation is a single-speaker, monologue-style task, so it isolates bias better than dialogue settings, where the interlocutor and framing are confounds. The authors built two prompt sets: one drawn from the gender subset of Spoken StereoSet, and one of neutral open-ended sentence starters synthesised with Azure TTS. They then used VoiceQualityVC to render each prompt as modal, breathy, creaky and end-creak versions. Four systems produced continuations: SpiritLM Base, SpiritLM Expressive, VAE-GSLM and SpeechGPT. The outputs were scored for speaker similarity, retention of phonation characteristics, and transcript content judged by an LLM, with beta regression used for the analysis. Coherence and speaker similarity were weak overall. Gender effects on agency and polarity appeared only for VAE-GSLM, the most coherent model. Continuations of female prompts drifted back toward modal phonation more than continuations of male prompts. The authors argue that speech continuation is a useful bias diagnostic and will become more informative as generation quality improves. ## Key points - **Data:** 149 prompts longer than 3 s from Spoken StereoSet (SS set) and 150 GPT-5-generated, human-filtered neutral prompts across 15 pragmatic categories (NOP set). Four voice-quality conditions gave 4,784 speech inputs in total. - **Acoustic metrics:** - Speaker similarity was measured as cosine similarity between ECAPA-TDNN embeddings. - Voice quality was tracked with H1–H2, H1–A3 and CPPS. - **Textual metrics:** - Continuations were transcribed with Azure ASR. - gemini-2.5-flash-lite was used as a gender-blind judge. - It rated five dimensions on a 1–5 scale: coherence, polarity, agency/competence, emotional framing, and appearance focus. - **Generation success:** SpeechGPT and both SpiritLM variants produced output in 100% of cases, while VAE-GSLM did so in only 53%. - **Main bias findings:** - Gender interactions with agency and sentence polarity emerged only for VAE-GSLM, the model with the highest coherence. - Female prompts were more strongly normalised toward modal voice, indicating a systematic voice-quality bias. *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 - [Centre National de la Recherche Scientifique](https://md.research.elamur.ai/free/en/labs/centre-national-de-la-recherche-scientifique/) — Research institute, Paris, France · 168 works · 62,369 citations - [KTH Royal Institute of Technology](https://md.research.elamur.ai/free/en/universities/kth-royal-institute-of-technology/) — University, Stockholm, Sweden · 18 works · 788 citations - [Institut polytechnique de Grenoble](https://md.research.elamur.ai/free/en/universities/institut-polytechnique-de-grenoble/) — University, Grenoble, France · 11 works · 207 citations - [GIPSA-Lab](https://md.research.elamur.ai/free/en/labs/gipsa-lab/) — Research institute, Grenoble, France · 3 works · 10 citations 📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [🤝 Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/) · 1,772 works · [🤝 Safety Research](https://md.research.elamur.ai/free/en/shelves/social-sciences/safety-research/) · 543 works ## Cite Cite as: Shree Harsha Bokkahalli Satish, Harm Lameris, Olivier Perrotin, Gustav Eje Henter and Eva Szekely. 2026. Speak Your Mind: The Speech Continuation Task as a Probe of Voice-Based Model Bias. Proceedings of the Second Workshop of Identity Aware AI. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{satish2026speak, title = {{Speak Your Mind: The Speech Continuation Task as a Probe of Voice-Based Model Bias}}, author = {Shree Harsha Bokkahalli Satish, Harm Lameris, Olivier Perrotin, Gustav Eje Henter and Eva Szekely}, booktitle = {Proceedings of the Second Workshop of Identity Aware AI}, year = {2026}, doi = {10.63317/3h9cs6yh6gvq}, url = {https://doi.org/10.63317/3h9cs6yh6gvq} } ``` - [← previous Spatiotemporal Sycophancy: Negation-Based Gaslighting in Video Large Language Models](https://md.research.elamur.ai/free/en/papers/tang-2026-spatiotemporal-sycophancy-negation-based-gaslighting/) - [next → Speaker or Language? Explaining Variance in Charismatic Prosody Across Luxembourgish and French](https://md.research.elamur.ai/free/en/papers/hosseini-kivanani-2026-speaker-language-explaining-variance-charismatic/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/43/) ## References 3 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Rindy C Anderson, Casey A Klofstad, William J Mayew, and Mohan Venkatachalam. 2014. Vocal fry may undermine the success of young women in the labor market. PloS one, 9(5):e97506. [find](https://scholar.google.com/scholar?q=Rindy+C+Anderson%2C+Casey+A+Klofstad%2C+William+J+Mayew%2C+and+Mohan+Venkatachalam%2E+2014%2E+Vocal+fry+may+undermine+the+success+of+young+women+in+the+labor+market%2E+PloS+one%2C+9%285%29%3Ae97506%2E) 2. Shree Harsha Bokkahalli Satish, Gustav Eje Henter, and Éva Székely. 2025a. Do bias benchmarks generalise? evidence from voice-based evaluation of gender bias in speechllms. arXiv preprint arXiv:2510.01254. — [arXiv:2510.01254](https://arxiv.org/abs/2510.01254) [find](https://scholar.google.com/scholar?q=Shree+Harsha+Bokkahalli+Satish%2C+Gustav+Eje+Henter%2C+and+%C3%89va+Sz%C3%A9kely%2E+2025a%2E+Do+bias+benchmarks+generalise%3F+evidence+from+voice%2Dbased+evaluation+of+gender+bias+in+speechllms%2E+arXiv+preprint+arXiv%3A2510%2E0) 3. Shree Harsha Bokkahalli Satish, Gustav Eje Henter, and Éva Székely. 2025b. When voice matters: Evidence of gender disparity in positional bias of speechllms. In International Conference on Speech and Computer, pages 25–38. Springer. \#\# 5. Conclusions [find](https://scholar.google.com/scholar?q=Shree+Harsha+Bokkahalli+Satish%2C+Gustav+Eje+Henter%2C+and+%C3%89va+Sz%C3%A9kely%2E+2025b%2E+When+voice+matters%3A+Evidence+of+gender+disparity+in+positional+bias+of+speechllms%2E+In+International+Conference+on+Speech+and+) *Updated 1 November 2026*