Skip to main content
An agent can use a cascaded voice pipeline or a realtime model, depending on the models available in your workspace.

Cascaded pipeline

A cascaded pipeline uses separate components for speech-to-text, language generation, and text-to-speech. It gives you direct control over the transcription model, LLM, and voice.

Realtime pipeline

A realtime pipeline uses a realtime model for the conversation. Some configurations use realtime audio directly, while others pair a realtime model with a selected text-to-speech voice.

Choose a configuration

Select the conversation language first, then choose compatible models and a voice. Review the model’s supported language, context window, and specialization in the builder. Use a representative simulated call to compare:
  • Transcription accuracy for names, numbers, and domain terms
  • Voice quality and pronunciation
  • Response latency
  • Interruption and endpointing behavior
  • Action execution and structured output
Change one model or speech setting at a time so you can attribute the result.