Topics we’re looking for.
SafeVoiceAgents welcomes work on — but not limited to — the following areas. In line with the exploratory nature of research workshops, we especially encourage unpublished, ongoing, or emerging work.
Voice-specific risks
Prosody, emotion, speaking style, interruption, overlap, ASR errors, latency, background audio, and full-duplex interaction.
AI safety methods for voice agents
Red teaming, jailbreaks, refusal, safety tuning, guardrails, and monitoring for spoken, streaming and agentic settings.
Safety and alignment of speech, audio and omni-modal foundation models
Why and where safety fails inside these models; alignment and safety training for speech and codec-token models where text guardrails and ASR mediation fall short; misuse of speech generation and voice cloning.
Safe action in voice-driven multimodal and embodied agents
Safety when spoken instructions drive tool use, device control, robots, vehicles and wearable assistants, including speaker verification, intent confirmation, and misheard or misattributed commands.
Safety evaluation and benchmarks
Metrics, datasets, taxonomies, judges, and human evaluation for multi-turn and real-time spoken dialogue.
Privacy, identity, and misuse
Voice identity exposure, always-on microphones, on-device processing, consent, data retention, impersonation, and deepfake-enabled misuse by or through agents.
Manipulation, overreliance and other harms to people
How voice similarity and cloning, tone, emotion and timing affect trust, disclosure, overreliance, persuasion, emotional dependence and wellbeing; measuring these effects and aligning agents not to exploit them.
Safety for vulnerable users and in high-stakes settings
Children, older adults, accessibility, education, physical and mental health.
Policy, deployment, and governance
Auditing, incident reporting, responsible release, and compliance for voice agents.
The workshop proposal is pending acceptance. Submissions are planned through OpenReview; the link will be posted here if accepted. All submissions are non-archival and will be public on OpenReview.
Full papers: up to 8 pages. Tiny papers: up to 4 pages, including late-breaking results, simple unpublished ideas, re-analyses, negative results and fresh perspectives. Position papers and demos are also welcome. Formatting details will follow acceptance.
Proposed important dates
These dates are proposed and subject to workshop acceptance.
- 3 December 2026
- Call for papers
- 25 January 2027
- Submission deadline
- 15 February 2027
- Reviews due
- 24 February 2027
- Author notification
- 26 March 2027
- Camera-ready papers and public schedule with talk titles
Submission and review policies
New work first. Unpublished and ongoing work is prioritized. Papers already accepted at a machine-learning conference, including the ICLR 2027 main track, may be presented as posters only and are not eligible for oral or spotlight slots. Authors must declare prior publication.
Double-blind review. Three reviews per paper, with at most three papers per reviewer. No organizer, meta-reviewer or reviewer handles a paper from their own organization or with a declared conflict; an unconflicted organizer handles assignment.
AI assistance. Authors may use AI assistance and must disclose it. AI systems cannot be authors or reviewers. Tiny papers must be primarily human-authored, following the ICLR tiny-paper policy.
Awards. Best Paper, Best Student Paper and Best Tiny Paper awards will be selected by the program committee independently of sponsors.