Voice agents

Enable customers to speak directly to an AI agent in their browser using real-time WebRTC audio.

Voice agents require a Deepgram API key in your settings. Deepgram handles speech-to-text and text-to-speech; you pay Deepgram directly.

How a voice call works

1

Customer clicks "Call" on your website

A WebRTC connection opens in-browser, requiring no app download. Audio streams directly from the microphone.

2

Speech is transcribed in real time

Deepgram Nova-3 converts spoken words to text with smart formatting, meaning numbers, times, and dates are structured automatically. Supports English, Spanish, French, German, Japanese, and more.

3

Agent queries your knowledge base & responds

The transcript passes through your agent's RAG pipeline. The answer is converted to speech using a Deepgram Aura voice (Asteria, Stella, or Orion) and streamed back.

Tuning response timing Admin

Under Agent Settings → Voice → Endpointing, set the silence threshold (in milliseconds). This controls how long the caller must pause before the agent considers them done speaking.

Lower values (e.g. 300ms) make the agent respond faster but risk cutting off slow speakers. Higher values (e.g. 700ms) feel more natural for conversational pauses.