Every turn must make clear whose floor it is

What it is

Navigation says “looking up a route,” then neither asks a question nor produces an audible close. People in the car cannot tell whether to speak a waypoint or stay quiet for the next sentence. Per-turn floor ownership requires that each beat make hearable who holds the right to speak, and whose opening is next legal. It governs conversational handoff, not the acoustic decision that an utterance has ended. When ownership is unclear, people barge in, wait through empty air, or treat the system’s thinking as already their turn.

Why it happens

Human talk runs on adjacency pairs: a question projects an answer, a greeting a greeting. While the second pair-part is empty, the floor defaults to the party that was asked. Systems break that projection in concrete ways: asking and then immediately reading the next paragraph, using a declarative while waiting for an answer, occupying the floor with “one moment” without saying how long the hold lasts. Users split strategy — some insert, some wait for a gap they believe they hear. Both strategies hitting the same state machine produce overlap and “I thought you were done.” Initiative in task dialogue is allowed to change hands, but the change must be hearable: a rising question, a named choice, or an explicit yield. Without those signals, the system’s internal listen/speak state and the state the user assumes fall out of alignment.

Studying it

Conversation analysis on live or Wizard-of-Oz audio: code each beat for holder, whether the other party accepted the handoff, overlap, incoming talk after a gap, and user insertion while the system is speaking. Independents: syntactic type of the system move (yes/no, open question, declarative-plus-wait) and whether a hearable hold marker follows the move. Dependent: mis-ownership rate — user speaks while the system still holds, or system takes back before the user has yielded.

Do not count every overlap as successful barge-in. One class of overlap is ownership failure: both parties think it is theirs. A lab that lets people “answer” with a button never sees floor fights, only choices.

Where it stops holding

One-way playback (news, a timer chime) never yields a floor, so ownership does not arise. Experts who already have a rhythm for a prompt set will take a turn even without a rise; new or occasional users of a skill do not have that rhythm. In multi-party scenes the floor also has to be allocated among people; marking “the user’s turn” does not mark which person. Reading ownership as “leave silence after the system speaks” tangles it with timeout policy — silence can be a yield or still thinking; ownership has to be marked in content and prosody, not by blankness alone.

Applying it

• At the end of every system utterance it must be answerable “who speaks next”: a question stops on the question; a hold gets a hold sound or word (“still looking”); a yield gets an audible completion (“found it. highways or fewer lights?”).
• Ban glued moves that ask and then keep talking. To keep talking, mark the hold first. Once the user speaks, treat it as a yield; do not drop that beat as noise and re-ask the same question.
• When initiative moves from system to user, build the second pair-part with a choice or a yes/no; do not wait in a declarative.
• How to check: code ten task recordings beat by beat. Simultaneous openings, or the user asking “should I talk now,” count as an ownership miss. Clusters on one skill: change the move type first, not the recognizer.

Related

• Same group: M1.03.1 Each extra turn stacks recognition and understanding error · M1.03.3 If the turn budget is spent, hand the task to another channel
• Nearby: M1.08 Turns and Floor Management · M1.12 Opening and Closing a Dialogue · M3.07 Rate, Pauses and Prosody
• Search terms: per-turn floor ownership · adjacency pair · initiative