Voice

An entity can speak its chat replies aloud. Speech runs locally on a device, through the desktop app's Voice Server add-on.

#What you need

• A device with the desktop app, online, with the entity deployed to it. See Desktop app setup.
• The Voice Server add-on installed from the app's Add-ons section. It needs an NVIDIA GPU with CUDA and at least 6 GB of GPU memory.
• For voice cloning, the Qwen TTS add-on, which needs at least 4 GB of GPU memory.

Voice is not part of MCP. Agents in Claude Code, ChatGPT, and other clients use the entity's memory, not its voice.

#Configure the entity's voice

1. Open the entity and go to Settings, then the Voice section.

2. Turn on Voice Enabled.

3. Choose a Provider and a Voice:

   Provider	Voices	Notes
   Piper (Local)	alba, marius, javert	Runs entirely on the device.
   OpenAI TTS	alloy, echo, fable, onyx, nova, shimmer	Uses OpenAI's speech service.
   ElevenLabs	Your choice	Paste the ElevenLabs Voice ID from your ElevenLabs dashboard.
   Qwen3 (Voice Cloning)	A custom voice	Upload a Voice Profile from the desktop app. The portal shows whether one is uploaded.

4. Click Save.

The device reads the configuration and speaks the entity's chat replies through the local Voice Server.

The Voice Server listens on the device at http://localhost:8765.

#Speaking to the entity

Spoken input, where you talk and the entity answers by voice, is in development. Today, type in Chat and the entity answers aloud.

#Troubleshooting

• The entity does not speak: check that Voice Enabled is on, the device is online, the entity is deployed to it, and the Voice Server add-on is running.
• The wrong voice plays: save the configuration again, and check that the selected provider's voice exists. ElevenLabs needs a valid Voice ID.
• No cloned voice: the Qwen3 provider needs a Voice Profile uploaded from the desktop app.