The path of one turn
Click a stage to see which settings group tunes it. Every knob in this guide lives at exactly one of these stages.
The ten parts
Pipeline mental model
The path of one turn, vocabulary, and which settings live at each stage.
TTS voice & model
Provider, model, voice, language, custom endpoints. Why on-prem overrides them.
Speaking plans (barge-in)
Start/stop plans, word thresholds, acknowledgement vs interruption phrases.
Normalization & substitutions
The six toggles, Turkish number handling, regex substitutions, background sound.
STT core
Provider, model, language constraint, temperature, transcription prompt.
Audio filters
AIC (dead), DeepFilterNet, WebRTC APM, smart turn. Order and cost.
VAD / endpointing
The densest part: confidence, volume, start/stop delay, timeout, echo grace.
Re-ask & keyterm bias
The dead re-ask group, plus keyterms, boost strength, anti-parrot.
Call-level controls
Silence timeout, max duration, voicemail, DTMF state machines.
Tuning playbooks
Scenario recipes, the mapping recap, and how to verify a change is live.
Defaults & ranges cheat-sheet
Every voice setting across Parts 2–9, in all three names: the dashboard label you click, the configKey you set via the API, and the runtime name inside pipecat-agent (often different — see the full map in NAMES.md). Filter by any of them; click a column to sort (Status groups dead → gotcha → live). Two defaults shown when defaults.ts and validators.ts disagree.
| Dashboard label | Config key | Runtime | Range | Default | Status |
|---|