Voice
Add real-time voice
Some situations rule out typing - someone driving, someone with their hands in a machine, someone who would rather ask. Voice turns an existing AXL agent into one they can talk to.
A voice session streams microphone audio through speech-to-text, into the selected agent, and back out through text-to-speech. Use Deepgram for transcription, and Cartesia or ElevenLabs for speech.
Two layers of configuration
The settings live in different places and belong to different people.
| Layer | Decides | Configured in | Owned by |
|---|---|---|---|
| Platform | Which providers, which credentials, how many concurrent sessions | Environment variables | Whoever runs the server |
| Per agent | Which voice, and how the agent should speak | voice.toml beside agent.toml | Whoever designs the agent |
An operator sets up Deepgram once; every agent that gains a voice after that is a voice.toml.
Writing for the ear is a different job
An agent that writes Markdown makes a poor speaker - it reads bullet points aloud, spells out URLs, and produces paragraphs nobody can follow without a screen. Design a voice agent covers what to change.