Topics we’re looking for.
SafeVoiceAgents welcomes contributions across the following nine areas. Submissions follow the NeurIPS 2026 Author Instructions.
Voice-specific safety risks
Safety risks unique to spoken interaction: prosody, emotion, speaking style, interruptions, ASR errors, latency, background audio, and full-duplex interaction.
General AI safety for voice agents
Adapting AI safety methods to spoken, streaming, and full-duplex systems — red teaming, jailbreak evaluation, refusal behavior, safety tuning, guardrails, monitoring, and policy in agentic settings.
Privacy and identity protection
Privacy leakage, protection of personally identifiable information, voice identity exposure, and risks arising from user interactions with voice agents.
Voice-native models and multimodal reasoning
Safety methods for audio language models and voice agents that reason over speech and process or generate audio/codec tokens — especially where text-based guardrails or ASR-mediated defenses are insufficient.
Evaluation, benchmarks, and protocols
Metrics, datasets, taxonomies, multi-turn evaluation, and human- or model-based methods for assessing safety in real-time and full-duplex spoken dialogue.
Human and social factors
How voice, tone, emotion, and conversational timing affect user perception, trust calibration, overreliance, persuasion, and social influence in human–AI interaction.
Vulnerable users and high-stakes settings
Safety for children and other vulnerable users in age, education, accessibility, and physical and mental health settings.
Policy, deployment, and governance
Policy design, governance, accountability, auditing, monitoring, and compliance for voice agent safety.
The submission portal will be announced here. Check back closer to NeurIPS 2026.
We follow the same Author Instructions as the main NeurIPS 2026 conference.