Skip to main content
Dead air is silence on a call that outlasts a threshold: neither the caller nor the agent speaking for longer than expected. attach() runs a per-session watchdog that polls for it and writes one event per silence stretch.

​
What counts as dead air

DeadAirDetector polls once a second (poll_interval_seconds, default 1.0) and asks an activity_probe callback for the session’s last-known activity timestamp. If the gap since that timestamp is at least threshold_seconds (default 3.0), it emits one DeadAirEvent and latches so it won’t fire again until activity resumes and the gap re-crosses the threshold: one event per silence stretch, not one per poll (src/voicegateway/middleware/dead_air_detector_middleware.py:17-18,43-124). The probe attach() supplies is driven by the same four speech-boundary events that feed turn capture. While either party is speaking the clock reports now, so a long utterance cannot trip a dead-air event. That is a guarantee of the wiring, not a tuning artifact. An event row carries: Rows land in dead_air_events and are read back oldest-first by session, or counted by session or time window.

​
Turning it on

On by default:
VOICEGW_DEAD_AIR=0 forces it off fleet-wide and beats the argument. It is a separate flag from turns even though both ride the same speech events, because dead air polls: one task per session, once a second, for the life of the call. That is a standing per-session cost, so it is switchable on its own.
LiveKit only. Pipecat accepts dead_air= for signature parity but has no equivalent speech-boundary events, so its dead-air view stays empty.
Events can also be pushed directly with POST /v1/ingest/dead-air; see the HTTP API.

​
Tuning the threshold

dead_air_threshold_seconds under projects.<id>.metrics in voicegw.yaml, default 3.0, per project rather than global. Three seconds is aggressive for an agent with a slow LLM: raise it if normal thinking time is firing events. The one-second poll interval is a constructor argument with no config key.

​
When the watcher doesn’t start

The watcher is scheduled with loop.create_task(...), which needs a running asyncio event loop when the session is wired up. Without one (synchronous code, a sync test rig) the auto-start is skipped: nothing crashes, no watcher runs, and the skip is logged once at debug.

​
Where you see it

GET /api/sessions/{id}/dead_air returns one session’s events, oldest first, with the same tenant-scoped 404 as the parent session (src/voicegateway/server/api/dashboard/sessions.py:225-248; full shape in the Dashboard API). The dashboard has no UI consumer for that raw list yet. An aggregate count across the filtered window (dead_air_event_count) surfaces via GET /api/metrics and is shown as a card on the dashboard’s Costs page, Conversation tab, alongside the response speed and talk-over rate covered under Turns and transcripts.