Turn-taking is the protocol by which two speakers in a conversation know whose turn it is to talk. Humans manage it with a combination of prosody, eye contact, intake breaths, syntactic completeness, and social context. In a voice-only AI conversation, all of those signals except the audio ones are gone.
A voice agent has to decide, turn by turn, when the caller is yielding the floor, when the agent should continue, when an interruption is real versus a backchannel ("uh-huh, yeah, okay"), and when to pause for thought versus push through. Each decision is the result of a model running on the audio stream, not a fixed rule.
The failure modes are uncanny in both directions. Too aggressive and the agent steamrolls the caller mid-sentence. Too passive and the agent waits awkwardly while the caller wonders if the line dropped. Tuning turn-taking is per-vertical work; calm clinical conversations have different rhythm than fast-paced auto-shop calls.
Vorel turn-taking is tuned per vertical. Clinic profiles give the caller more space; dispatch profiles for home services keep the back-and-forth tight to match how the customer expects the call to go.