Turns and interruptions

Turns and epochs

Every user turn gets a number, its epoch. The turn driver is the only component that advances it. Bot-side frames (LLM text, TTS audio, playback marks) carry the epoch they were produced for; every stage on the bot side drops frames from a stale epoch. This is the one mechanism that keeps a late audio chunk, a late LLM token or a late tool result from leaking into the next turn.

Deciding that the caller has finished

A pure state machine (TurnMachine) with injected time decides when a turn ends. The policies:

• Adaptive (default). When the VAD sees the caller fall silent, two things start at once: the STT Finalize, and the Smart Turn model (an on-device end-of-turn classifier, run on the inference pool). If the model says the utterance is complete, the turn ends as soon as the STT final has arrived; if it says the caller is not done, the turn waits, with a fallback deadline (fallback_silence_ms) so an undecided pause cannot hang the call.
• Fixed. The turn ends after a fixed silence.

Because the turn end waits for the STT final (a "finals gate", bounded by a timeout), the added wait is the larger of the two, not their sum.

With speculation enabled, the LLM request is started at a candidate end of turn and committed only when the turn really ends; a read-only tool may run on a speculation, a side-effect tool is held until the commit. What this buys, and what it costs in extra requests, is measured in bench/results/speculation-fixtures.md (summarised in Benchmark results).

Barge-in

When the caller starts speaking while the bot is speaking, the turn driver judges the candidate interruption (the interruption policy: minimum words, backchannel lists such as "yeah" or "uh-huh" per language, an echo guard) and, if it is real:

1. starts the new epoch (an Interrupt always precedes the new UserTurnStarted);
2. sends Clear straight to the transport on a fast path, so queued audio is dropped and the client is told to flush what it buffered;
3. cancels the LLM round, the segmenter and the open TTS contexts;
4. cuts the context to what was actually played (Played-text truncation).

The budget for step 2, decision to Clear applied, is 1 ms at p95. After it, no stale audio reaches the wire.

Read more: "Turn-taking" in Architecture (the state diagram and the policies), sections 2 and 3 of the pipeline design (epoch ownership and the barge-in path step by step), and the interruption keys in Configuration.