Everything Counts: The Managed Omnirelevance of Speech in Human-Voice Agent Interaction
For researchers in human-computer interaction, this work highlights a persistent interactional challenge in voice agent design that has been underexplored.
This study examines how humans manage the risk that any speech may trigger an unwanted response from voice agents, documenting practices across rule-based and modern systems. The omnirelevance of human speech is identified as a constitutive feature of human-agent interaction.
To this day, turn-taking models determining voice agents' conduct have been examined primarily from a technical point of view, while the ways in which they emerge as interactional constraints or resources for human conversationalists in situ remain underexplored. Drawing on a detailed analysis of corpora of naturalistic data, we document how humans' conduct was produced in reference to the ever-present risk that, each time they spoke, their talk might trigger a new uncalled-for contribution from the artificial agent. We examine this phenomenon in interactions involving rule-based robots from a 'pre-LLM era' as well as the most recent voice agents. This 'omnirelevance of human speech' (i.e., the possibility that a conversational agent may erroneously respond to any speech it detects) emerged as a constitutive feature of these human-agent encounters. We describe some of the practices through which humans managed these artificial agents' turn-taking conduct. Given recent improvements in voice capture technology, we ask whether this 'omnirelevance of human speech' weighs even more heavily on human practices today than in the past.