What is Turn?
Also called Conversational Turn, Dialogue Turn.
A turn is one contribution to a conversation by one party, followed by the other party's opportunity to respond. In text systems a turn is usually a single message, while in speech it is a stretch of talk bounded by the speaker yielding the floor. Counting turns is how dialogue systems measure conversation length and structure context.
For model driven assistants, the turn is the unit of context. History is assembled as alternating user and assistant turns, sometimes with tool calls and their results interleaved, and the model sees that sequence when producing the next reply. Multi turn simply means the system carries information across turns rather than treating every message as independent.
Speech makes turn boundaries genuinely hard. People pause mid thought, trail off, and interrupt, so a voice system needs endpointing, the decision about when a caller has actually finished rather than merely paused. Waiting too long feels unresponsive, and cutting in too early talks over the caller. Human conversation switches speakers within a few hundred milliseconds, which sets an uncomfortable bar.
Turn counts are a common but blunt metric. Fewer turns can mean an efficient resolution or a person giving up, and more turns can mean careful clarification or a system that keeps failing to understand. Turns to resolution paired with an outcome measure is far more informative than either number alone, and the same applies to average handling time.
Turn structure also shapes cost and latency. Every turn in a model driven system re-sends accumulated context, so long conversations get slower and more expensive as they go, which is one practical reason for summarizing older turns rather than carrying every word forward indefinitely.
Key points
- One contribution by one party, then the other's opportunity
- The assembly unit for model context in multi turn systems
- Voice needs endpointing to detect when a speaker has finished
- Turn count alone does not indicate success or failure
- Each turn re-sends context, so long conversations cost more
In practice
A caller asks a voice agent to change a flight. Turn one is the request, turn two the agent asking which booking, turn three the reference number, turn four two offered alternatives. The exchange resolves in six turns. A second caller also finishes in six turns, but their sixth is 'forget it, I will call back later,' which is exactly why turn count needs an outcome beside it.