SafeVoiceAgents

Voice is becoming the interface that connects people with AI. Digital assistants in tutoring and medicine, and embodied agents in cars and robots, are moving from cascaded ASR–LLM–TTS pipelines to speech-native, full-duplex models that listen while they speak. Voice makes interaction possible for people who cannot read, type or see a screen, and in situations where hands and eyes are occupied.

SafeVoiceAgents connects AI safety and alignment, speech and audio, agents and embodied AI, security, privacy and human-centered AI. We bring these communities together to develop speech-native threat models, real-time safeguards, shared evaluation principles and a research agenda for trustworthy voice agents.

Voice agents inherit the risks of LLM agents, amplify them, and introduce new ones. An agent can book, pay, send messages and control homes, cars or robots. Its safety depends not only on what is said, but how it is said, who is speaking, and whether a command is correctly understood before it becomes action.

Failures arise at four levels: audio foundation models, spoken interaction, agent systems and evaluation. Text-learned safety does not reliably transfer to audio. Prosody, emotion, accent and background sound can carry threats beyond the transcript; interruption and turn-taking create real-time attack surfaces. Voice conveys biometric and emotional data, can persuade or encourage overreliance, and leaves bystanders exposed to always-on microphones. Safety must also hold for different languages, accents, ages and abilities.