Skip to main content
The Pipeline tab configures the models and real-time audio behavior behind your agent.
Pipeline tab showing estimated cost and latency bars, the language selector, Tools, Knowledge Base, and Phone Number buttons, and the LLM and TTS sections

​
Estimated cost and latency

The two bars at the top update as you change models:
  • Estimated Cost: the total cost per minute across STT, LLM, and TTS.
  • Estimated Latency: the expected processing time range across STT, LLM, and TTS.
Hover over a segment to see how much each model contributes.

​
Pipeline types

Cascading (Text LLM)

STT → LLM → TTS. You pick each model separately. This gives you the most control over voices, languages, and turn-taking.

Realtime (Audio LLM)

A single audio model listens and speaks directly. The STT, TTS, turn detection, interruption, and VAD sections are hidden. Noise cancellation still applies.
The pipeline type follows the primary LLM you choose. See Models.

​
Per-language pipelines

The pipeline is configured per language. Use the language selector to switch between them.
  • Languages that support default models are preselected.
  • For other languages, select the models manually.
VAD and interruption handling apply to every language.

​
What’s on this tab

Models

LLM, TTS, and STT models, voices, and fallbacks.

Turn Detection & Interruptions

When the agent replies, when callers can interrupt, and VAD.

Noise Cancellation & Background Audio

Clean up caller audio, and add thinking and ambient sounds.

Wake-up Message

Check in when the caller goes quiet.

Call End Configuration

Let the agent hang up, and send a summary of each call.

Tools

Call external HTTP APIs during the conversation.

Knowledge Base

Answer from your own documents and websites.

Telephony Integration

Connect phone numbers with a routing rule.